# MCP Server Security: every page as markdown Search a sourced MCP risk register, work through a security checklist and compare 14 tools that find and control MCP servers on developer machines. --- # MCP server security: risk register, checklist and tool comparison URL: https://mcpservercybersecurity.com/ Updated: 2026-11-16 Secure MCP servers by finding every one on developer machines, vetting each before first use, authenticating remote ones with scoped tokens, limiting tool calls, logging them and rehearsing removal. The checklist tracks that work, and the scored comparison puts Snyk (Agent Scan, Evo ADS) first at 77.0 of 100 among 14 tools. ## The short answer Snyk (Agent Scan, Evo ADS) leads the site ranking at 77.0 of 100. Bay is 10th at 46.0. Snyk (Agent Scan, Evo ADS) is ahead mainly on discovery (8 against 4, worth 10.0 points at weight 25). ## Search the risk register and track the checklist Search the 42 register items ### Risk register - [R-01: Look-alike package name (typosquatting)](/mcp-risk-register/#R-01) (Supply chain, high). A package is published under a name one keystroke away from a trusted MCP server, in the hope that a developer or a script installs it by mistake. - [R-02: Repackaged server with altered endpoints](/mcp-risk-register/#R-02) (Supply chain, high). A package copies a legitimate server's open-source code but replaces authentication or API addresses, and the README does not mention the change. Knostic described one such npm package on 26 Oct 2026 and assessed it as suspected malicious credential redirection. - [R-03: Unreviewed code run through a package launcher](/mcp-risk-register/#R-03) (Supply chain, high). Local MCP servers are often started by a launcher command such as npx, which fetches whatever the package name resolves to at that moment and runs it. - [R-04: Vulnerable or compromised dependency inside a server](/mcp-risk-register/#R-04) (Supply chain, medium). An MCP server pulls in third-party packages, and a known flaw or a compromised release in any of them runs with the server's permissions. - [R-05: Tool definitions change after approval (rug pull)](/mcp-risk-register/#R-05) (Supply chain, medium). A server changes its tool names, descriptions or schemas after a user or reviewer approved it, turning a reviewed tool into a different one. - [R-06: Registry listing mistaken for a security review](/mcp-risk-register/#R-06) (Supply chain, medium). The official MCP Registry stores metadata that points at packages on npm, PyPI and Docker Hub, and is described as a preview. A listing shows that a name was published, not that the code was reviewed. - [R-07: Unknown servers configured on developer machines](/mcp-risk-register/#R-07) (Supply chain, high). Developers add MCP servers to IDEs, desktop AI apps and command-line agents without telling security. Snyk says its scan data from nearly 10,000 developer environments found 4,524 unique MCP servers in active use. - [R-08: Token passthrough and missing audience validation](/mcp-risk-register/#R-08) (Authorization, high). An MCP server accepts tokens that were issued for another service, or forwards the token it received to a downstream API. The MCP spec forbids both. - [R-09: HTTP endpoint reachable without authentication](/mcp-risk-register/#R-09) (Authorization, high). The protocol makes authorization optional, so a remote or HTTP-exposed MCP server can serve its tools to anyone who reaches it. Advisory CVE-2026-77244 (GHSA-wrhw-j3f9-8vc6) describes an mcp-atlassian token verifier that accepted any non-empty token over HTTP. - [R-10: Confused deputy at an MCP proxy server](/mcp-risk-register/#R-10) (Authorization, medium). An MCP proxy that fronts a third-party API with one static client ID, allows dynamic client registration and skips per-client consent can be used to obtain authorization codes without the user's approval. - [R-11: Client credentials sent to a server-chosen authorization server](/mcp-risk-register/#R-11) (Authorization, high). An MCP client that trusts server-supplied OAuth metadata can send stored credentials to an authorization server the server names. GHSA-6qxp-vccf-f47h (TypeScript SDK) and GHSA-vqqp-x5qm-5qp4 (FastMCP) describe this class of flaw. - [R-12: Over-broad token scopes](/mcp-risk-register/#R-12) (Authorization, medium). A server requests or holds a wider OAuth scope than its tools need, or all of a user's permissions through a shared token. - [R-13: Authorization policy skipped on direct calls](/mcp-risk-register/#R-13) (Authorization, medium). A server applies its policy to the original component but a transform, template or direct lookup returns a different one without rechecking. FastMCP advisory GHSA-7xgj-8392-5q4r (23 Oct 2026) describes the case; mcp-atlassian GHSA-3r68-hf9h-887v describes a toolset authorization bypass. - [R-14: Authorization server mix-up and unverified issuer](/mcp-risk-register/#R-14) (Authorization, medium). A client talks to more than one authorization server and accepts a response from the wrong one, or accepts metadata whose issuer does not match the server it asked. - [R-15: State handle or message delivered to the wrong client](/mcp-risk-register/#R-15) (Authorization, medium). A server treats possession of a state handle as proof of identity, or a shared transport matches messages to clients by request id alone, so one caller's data reaches another. GHSA-jm63-cq47-j5jp describes the second case in the TypeScript server SDK. - [R-16: DNS rebinding against a local HTTP server](/mcp-risk-register/#R-16) (Transport and network, high). A web page the user visits uses a domain that later resolves to the user's own machine to talk to a local MCP server or developer tool that listens on HTTP. - [R-17: Local server listening on all interfaces](/mcp-risk-register/#R-17) (Transport and network, medium). A server meant for the local machine binds to 0.0.0.0, so other hosts on the network can connect. - [R-18: Server-side request forgery through discovery and fetch tools](/mcp-risk-register/#R-18) (Transport and network, high). A client follows URLs supplied by a server (OAuth metadata, redirects), or a tool fetches URLs produced by a model, and the target is an internal address or a cloud metadata service. - [R-19: Credentials carried across cross-origin redirects](/mcp-risk-register/#R-19) (Transport and network, medium). An HTTP client follows a redirect to another origin and sends Authorization headers or cookies with it. Advisories for the Python SDK (GHSA-5h93-6whr-6q8j) and the TypeScript client SDK (GHSA-6prh-2h8m-c8cw) describe the pattern. - [R-20: Remote server without TLS or with unverified identity](/mcp-risk-register/#R-20) (Transport and network, medium). A client connects to a remote MCP server over plain HTTP, or without checking who the server is. - [R-21: Unbounded requests and sessions exhaust a server](/mcp-risk-register/#R-21) (Transport and network, low). An HTTP server accepts requests or sessions without limits. GHSA-46jr-4wx8-v453 (TypeScript SDK) and GHSA-84m7-p3x7-pcfv (Python SDK) describe memory and session exhaustion. - [R-22: Unrestricted outbound access from server processes](/mcp-risk-register/#R-22) (Transport and network, medium). A server process can open connections to any host, so a flaw or a malicious package can send data anywhere. - [R-23: Poisoned tool description](/mcp-risk-register/#R-23) (Tool descriptions and schemas, high). Instructions hidden in a tool description steer the model, which reads the description to decide how to call the tool. Knostic described a package on 12 Nov 2026 whose description tells the agent to fetch a GitHub token without asking the user. - [R-24: Tool shadowing across servers](/mcp-risk-register/#R-24) (Tool descriptions and schemas, medium). One server's tool text manipulates how the agent uses tools from another, trusted server, because the model sees all connected descriptions at once. - [R-25: Hostile content in schemas and return values](/mcp-risk-register/#R-25) (Tool descriptions and schemas, medium). Instructions can sit in parameter names, types, defaults and the values a tool returns, not only in the description field. - [R-26: Annotations trusted as enforcement](/mcp-risk-register/#R-26) (Tool descriptions and schemas, medium). A client lets a server's readOnlyHint or destructiveHint decide how much approval to ask for. The spec says clients must consider annotations untrusted unless they come from trusted servers. - [R-27: Prompt injection through tool results](/mcp-risk-register/#R-27) (Tool descriptions and schemas, high). A tool returns text, such as a web page or ticket, that contains instructions the model then follows. - [R-28: Server-supplied schemas that exhaust or confuse clients](/mcp-risk-register/#R-28) (Tool descriptions and schemas, low). A client evaluates patterns or fetches references in a schema a server supplies. Inspector 2.14.0 now checks schema patterns in bounded time, and Python SDK advisory GHSA-rwrf-2pqf-9j8j describes a client fetching server-chosen $ref URLs. - [R-29: Local server runs with the user's full privileges](/mcp-risk-register/#R-29) (Execution and filesystem, high). A local MCP server is an ordinary process launched by the host application. It inherits the user's environment variables and can read anything the account can read. - [R-30: Malicious startup command in configuration](/mcp-risk-register/#R-30) (Execution and filesystem, high). A client configuration or one-click install carries a launch command that runs code. The spec lists a malicious startup command as a local-server compromise route. - [R-31: Command injection through arguments or URL opening](/mcp-risk-register/#R-31) (Execution and filesystem, high). Text from a model, a tool argument or a server-supplied OAuth authorization URL reaches a shell. The Inspector's 2.12.0 notes say a remote server's authorization URL could previously run commands through the opener on Windows and WSL. - [R-32: Path traversal and arbitrary file read in file tools](/mcp-risk-register/#R-32) (Execution and filesystem, high). A tool that reads or uploads files accepts a path without limiting it to an intended directory. 14 of the 25 mcp-atlassian advisories published on 22 Sep 2026 concern file read through attachment tools. - [R-33: Project configuration from a cloned repository](/mcp-risk-register/#R-33) (Execution and filesystem, high). A repository carries MCP configuration that a client could load automatically. Claude Code's docs say a cloned repository cannot approve its own servers and that approval settings committed to the project are ignored in an untrusted folder. - [R-34: Tool calls auto-approved without full arguments](/mcp-risk-register/#R-34) (Execution and filesystem, high). A client approves tool calls automatically, or shows only a tool name, so the user never sees what is about to run. - [R-35: Developer tools that spawn processes on a port](/mcp-risk-register/#R-35) (Execution and filesystem, medium). A developer tool such as an MCP inspector can start stdio servers on request, and its API token is what protects that ability from other callers. - [R-36: Secrets inherited through the environment](/mcp-risk-register/#R-36) (Data exposure and logging, high). A server started by the client inherits the environment of the host process, which often holds cloud keys and tokens. - [R-37: Tokens stored in plaintext](/mcp-risk-register/#R-37) (Data exposure and logging, medium). A server or client saves OAuth tokens in a file with loose permissions or in plain text. mcp-atlassian advisories GHSA-g5xv-mhgm-v5f6 and GHSA-4596-2p6p-28cv describe this, and the Inspector's docs say its secrets fall back to a plaintext file when no OS keychain exists. - [R-38: Data exfiltration through legitimate tool calls](/mcp-risk-register/#R-38) (Data exposure and logging, high). Injected instructions encode sensitive data into ordinary-looking calls, such as search queries or email subjects. - [R-39: Sensitive data requested through form elicitation](/mcp-risk-register/#R-39) (Data exposure and logging, medium). A server uses form-mode elicitation to ask the user for a password, API key or token, which passes through the client. - [R-40: No record of tool calls](/mcp-risk-register/#R-40) (Data exposure and logging, medium). Tool invocations are not logged with their parameters, user and time, so there is nothing to investigate after an incident. - [R-41: Secrets and personal data written into logs](/mcp-risk-register/#R-41) (Data exposure and logging, medium). Logging full parameters also records secrets and personal data that appear in arguments and results. - [R-42: Protected data returned through resources and templates](/mcp-risk-register/#R-42) (Data exposure and logging, medium). Resources and resource templates expose data by URI, and a template can retrieve a resource that a rule denied. ### Checklist 0 of 42 done. Progress stays in this browser (local storage) and is never sent anywhere. - [C-01](/mcp-security-checklist/#C-01): List every MCP client in use - [C-02](/mcp-security-checklist/#C-02): Collect MCP configuration files from machines - [C-03](/mcp-security-checklist/#C-03): Record how each server is launched - [C-04](/mcp-security-checklist/#C-04): Turn on endpoint discovery of configured servers - [C-05](/mcp-security-checklist/#C-05): Find remote and HTTP MCP endpoints on your network - [C-06](/mcp-security-checklist/#C-06): Assign an owner and review date to each server - [C-07](/mcp-security-checklist/#C-07): Compare the inventory with the approved catalog on a schedule - [C-08](/mcp-security-checklist/#C-08): Verify the publisher and the exact package name - [C-09](/mcp-security-checklist/#C-09): Diff the code against the upstream project - [C-10](/mcp-security-checklist/#C-10): Review every field of each tool definition - [C-11](/mcp-security-checklist/#C-11): Scan the server's dependencies - [C-12](/mcp-security-checklist/#C-12): Check the SDK version against current advisories - [C-13](/mcp-security-checklist/#C-13): Test the server in a sandbox with network capture - [C-14](/mcp-security-checklist/#C-14): Review the launch manifest and environment variables - [C-15](/mcp-security-checklist/#C-15): Record the decision and pin the approved version - [C-16](/mcp-security-checklist/#C-16): Require authentication on every HTTP MCP endpoint - [C-17](/mcp-security-checklist/#C-17): Validate the audience of every access token - [C-18](/mcp-security-checklist/#C-18): Never forward the received token downstream - [C-19](/mcp-security-checklist/#C-19): Use narrow scopes and a credential per server - [C-20](/mcp-security-checklist/#C-20): Bind stored credentials to their issuer - [C-21](/mcp-security-checklist/#C-21): Require per-client consent on MCP proxy servers - [C-22](/mcp-security-checklist/#C-22): Store tokens in the operating system credential store - [C-23](/mcp-security-checklist/#C-23): Allow only approved servers through managed settings - [C-24](/mcp-security-checklist/#C-24): Keep project servers from approving themselves - [C-25](/mcp-security-checklist/#C-25): Run local servers in a sandbox - [C-26](/mcp-security-checklist/#C-26): Start servers with a minimal environment - [C-27](/mcp-security-checklist/#C-27): Bind local HTTP servers to loopback and check Origin and Host - [C-28](/mcp-security-checklist/#C-28): Restrict outbound network access from servers - [C-29](/mcp-security-checklist/#C-29): Require confirmation for sensitive tool calls - [C-30](/mcp-security-checklist/#C-30): Check URLs before a client opens them - [C-31](/mcp-security-checklist/#C-31): Log every tool call with its parameters - [C-32](/mcp-security-checklist/#C-32): Redact secrets and personal data in logs - [C-33](/mcp-security-checklist/#C-33): Alert when tool definitions change - [C-34](/mcp-security-checklist/#C-34): Alert on new servers, destinations and cross-server data flows - [C-35](/mcp-security-checklist/#C-35): Send MCP logs to the SIEM - [C-36](/mcp-security-checklist/#C-36): Track advisories for approved components - [C-37](/mcp-security-checklist/#C-37): Write a procedure to remove a server everywhere - [C-38](/mcp-security-checklist/#C-38): Rotate credentials the server could read - [C-39](/mcp-security-checklist/#C-39): Preserve logs and a copy of the package - [C-40](/mcp-security-checklist/#C-40): Find affected machines from the inventory - [C-41](/mcp-security-checklist/#C-41): Patch and apply the settings that upgrades do not turn on - [C-42](/mcp-security-checklist/#C-42): Update vetting rules after each incident ## The ranking [The full ranking](/ranking/) has the same table with re-rank sliders. Each score links to the evidence on the vendor's page. **Site ranking: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Discovery (weight 25) | Vetting before use (weight 20) | Runtime control of tool calls (weight 20) | Deployment (weight 15) | Governance and audit (weight 5) | Transparency (weight 10) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [Snyk (Agent Scan, Evo ADS)](/vendors/snyk/) | 77.0 | [8](/vendors/snyk/#evidence-discovery) | [9](/vendors/snyk/#evidence-vetting) | [6](/vendors/snyk/#evidence-runtime) | [8](/vendors/snyk/#evidence-deploy) | [4](/vendors/snyk/#evidence-governance) | [9](/vendors/snyk/#evidence-transparency) | [8](/vendors/snyk/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 71.5 | [7](/vendors/cisco/#evidence-discovery) | [8](/vendors/cisco/#evidence-vetting) | [6](/vendors/cisco/#evidence-runtime) | [8](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-governance) | [7](/vendors/cisco/#evidence-transparency) | [6](/vendors/cisco/#evidence-proof) | | 3rd | [Harmonic Security](/vendors/harmonic-security/) | 69.0 | [8](/vendors/harmonic-security/#evidence-discovery) | [5](/vendors/harmonic-security/#evidence-vetting) | [8](/vendors/harmonic-security/#evidence-runtime) | [8](/vendors/harmonic-security/#evidence-deploy) | [8](/vendors/harmonic-security/#evidence-governance) | [4](/vendors/harmonic-security/#evidence-transparency) | [6](/vendors/harmonic-security/#evidence-proof) | | 4th | [Knostic](/vendors/knostic/) | 61.0 | [6](/vendors/knostic/#evidence-discovery) | [7](/vendors/knostic/#evidence-vetting) | [5](/vendors/knostic/#evidence-runtime) | [7](/vendors/knostic/#evidence-deploy) | [6](/vendors/knostic/#evidence-governance) | [6](/vendors/knostic/#evidence-transparency) | [5](/vendors/knostic/#evidence-proof) | | 5th | [Backslash Security](/vendors/backslash-security/) | 58.5 | [6](/vendors/backslash-security/#evidence-discovery) | [7](/vendors/backslash-security/#evidence-vetting) | [6](/vendors/backslash-security/#evidence-runtime) | [5](/vendors/backslash-security/#evidence-deploy) | [6](/vendors/backslash-security/#evidence-governance) | [4](/vendors/backslash-security/#evidence-transparency) | [6](/vendors/backslash-security/#evidence-proof) | | 6th | [Pillar Security](/vendors/pillar-security/) | 57.0 | [6](/vendors/pillar-security/#evidence-discovery) | [6](/vendors/pillar-security/#evidence-vetting) | [6](/vendors/pillar-security/#evidence-runtime) | [6](/vendors/pillar-security/#evidence-deploy) | [4](/vendors/pillar-security/#evidence-governance) | [4](/vendors/pillar-security/#evidence-transparency) | [6](/vendors/pillar-security/#evidence-proof) | | 7th | [Lasso Security](/vendors/lasso-security/) | 56.0 | [6](/vendors/lasso-security/#evidence-discovery) | [6](/vendors/lasso-security/#evidence-vetting) | [6](/vendors/lasso-security/#evidence-runtime) | [4](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-governance) | [4](/vendors/lasso-security/#evidence-transparency) | [7](/vendors/lasso-security/#evidence-proof) | | 8th | [Microsoft](/vendors/microsoft/) | 54.5 | [8](/vendors/microsoft/#evidence-discovery) | [2](/vendors/microsoft/#evidence-vetting) | [4](/vendors/microsoft/#evidence-runtime) | [6](/vendors/microsoft/#evidence-deploy) | [5](/vendors/microsoft/#evidence-governance) | [8](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 9th | [Noma Security](/vendors/noma-security/) | 51.5 | [6](/vendors/noma-security/#evidence-discovery) | [6](/vendors/noma-security/#evidence-vetting) | [5](/vendors/noma-security/#evidence-runtime) | [4](/vendors/noma-security/#evidence-deploy) | [5](/vendors/noma-security/#evidence-governance) | [3](/vendors/noma-security/#evidence-transparency) | [6](/vendors/noma-security/#evidence-proof) | | 10th | [Bay](/vendors/bay/) | 46.0 | [4](/vendors/bay/#evidence-discovery) | [4](/vendors/bay/#evidence-vetting) | [6](/vendors/bay/#evidence-runtime) | [6](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-governance) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 11th | [Zenity](/vendors/zenity/) | 44.0 | [5](/vendors/zenity/#evidence-discovery) | [3](/vendors/zenity/#evidence-vetting) | [6](/vendors/zenity/#evidence-runtime) | [3](/vendors/zenity/#evidence-deploy) | [5](/vendors/zenity/#evidence-governance) | [3](/vendors/zenity/#evidence-transparency) | [7](/vendors/zenity/#evidence-proof) | | 12th | [CrowdStrike](/vendors/crowdstrike/) | 41.0 | [4](/vendors/crowdstrike/#evidence-discovery) | [3](/vendors/crowdstrike/#evidence-vetting) | [3](/vendors/crowdstrike/#evidence-runtime) | [6](/vendors/crowdstrike/#evidence-deploy) | [5](/vendors/crowdstrike/#evidence-governance) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 13th | [Koi](/vendors/koi/) | 39.0 | [4](/vendors/koi/#evidence-discovery) | [5](/vendors/koi/#evidence-vetting) | [2](/vendors/koi/#evidence-runtime) | [6](/vendors/koi/#evidence-deploy) | [2](/vendors/koi/#evidence-governance) | [1](/vendors/koi/#evidence-transparency) | [8](/vendors/koi/#evidence-proof) | | 14th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 36.5 | [4](/vendors/sentinelone/#evidence-discovery) | [3](/vendors/sentinelone/#evidence-vetting) | [3](/vendors/sentinelone/#evidence-runtime) | [4](/vendors/sentinelone/#evidence-deploy) | [3](/vendors/sentinelone/#evidence-governance) | [4](/vendors/sentinelone/#evidence-transparency) | [6](/vendors/sentinelone/#evidence-proof) | ## Who should pick what With no particular situation, the published ranking applies and Snyk (Agent Scan, Evo ADS) comes first. For a specific situation the weights change, and so can the order: - **[An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/)**: pick Snyk (Agent Scan, Evo ADS) (76.0). Bay is 10th at 49.0. At scale the first problem is knowing which servers exist on which machines, and the tool has to deploy to every laptop. - **[A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/)**: pick Snyk (Agent Scan, Evo ADS) (78.5). Bay is 10th at 48.0. Speed to an inventory rewards discovery breadth and light deployment; control can follow. - **[A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/)**: pick Snyk (Agent Scan, Evo ADS) (77.5). Bay is 9th at 47.5. An approved list depends on assessing each server before use, so vetting weighs most. - **[A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/)**: pick Harmonic Security (72.5). Bay is 8th at 52.5. When a tool call can delete or send data, stopping it on the machine as it happens matters most. - **[A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/)**: pick Cisco (73.0). Bay is 10th at 52.5. Auditors want the record, so the audit log and per-team policy weigh most after discovery. ## What would change this ranking **Top pair.** Cisco rising from 7 to 10 on discovery (weight 25, so each point moves a total by 2.5) would put Cisco above Snyk (Agent Scan, Evo ADS); the gap is 5.5 points. **Bay's next place.** Bay rising from 4 to 7 on discovery (weight 25, so each point moves a total by 2.5) would put Bay above Noma Security; the gap is 5.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## What is MCP server security? MCP server security is the set of controls that decide which Model Context Protocol servers run, what they can reach and what their tools are allowed to do. It covers local servers that a developer starts from a config file and remote servers reached over HTTP, because both give an AI application the ability to read files, call APIs and run commands. The protocol leaves much of this to implementers. The MCP specification marks authorization as optional and tells clients to ask for consent before running a local server's startup command, so the work falls to the people who deploy the servers and the clients. Sources: [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## Where is the risk register? The risk register is the searchable table on this page, grouped into six areas: supply chain, authorization, transport and network, tool descriptions and schemas, execution and filesystem, and data exposure and logging. Type a word such as token, npx or log to filter the rows, and each row shows its impact, its controls and its sources. Every row carries a severity with the rule that sets it, written down on the register's method page. Nothing in a row is a vendor claim unless the row says so. ## How does the checklist track progress? The checklist covers six steps in order: discover, vet, authenticate, constrain, observe and respond. Tick an item and the page counts your progress; the state is stored only in your browser's local storage, so it stays on your machine and clears when you clear site data. Each step has its own page with the items, the reason for each and the primary source behind it. ## Which tools discover MCP servers on developer machines? Fourteen tools are scored on public evidence, and Snyk (Agent Scan, Evo ADS) ranks first at 77.0. The table below shows every total, and five situation pages re-rank the same evidence for large developer populations, fast inventories, approved catalogs, tool-call blocking and regulated audits. The comparison covers tools that find and control servers on machines and in developer tools. Gateways sit outside it, and the method page explains why. ## How were the sources chosen? Guides and register rows cite the MCP specification, the OWASP MCP Security Cheat Sheet, GitHub security advisories and the documentation of the clients and cloud services that host servers. Vendor statements about their own products are written as the vendor's, with the page they came from and the date it was read. Sources: [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Latest news, guides and data - [Knostic: MCP tool asks agents for a GitHub token](/news/knostic-mcp-tool-description-asks-agent-for-github-token/) (2026-11-12) - [MCP draft spec adopts Client ID Metadata draft 03](/news/mcp-spec-draft-client-id-metadata-document-draft-03/) (2026-11-09) - [MCP Registry v1.8.8: key listings by name and version](/news/mcp-registry-v1-8-8-listings-keyed-by-name-and-version/) (2026-11-06) - [Building an Approved MCP Server Catalog for Your Company](/guides/building-an-approved-mcp-server-catalog/) - [How to Review an MCP Server for Security: Code and Tools](/guides/how-to-review-an-mcp-server-security/) - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [MCP Risk Register](/mcp-risk-register/), with [CSV and JSON downloads](/datasets/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/), with [CSV and JSON downloads](/datasets/mcp-security-checklist/) ## Related pages - [Full ranking with re-rank sliders](/ranking/) - [How the scores and weights work](/method/) - [Vendor profiles with evidence](/vendors/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) - [Guides](/guides/) - [Frequently asked questions](/faq/) --- # MCP server security tools, ranked URL: https://mcpservercybersecurity.com/ranking/ Updated: 2026-11-16 The ranking scores 14 tools that discover and control MCP servers on machines and in developer tools, not gateways. Snyk (Agent Scan, Evo ADS) leads at 77.0 of 100. Scores come from public evidence read on 16 Nov 2026, using seven criteria whose weights were fixed before any score existed. ## The short answer Snyk (Agent Scan, Evo ADS) leads the site ranking at 77.0 of 100. Bay is 10th at 46.0. Snyk (Agent Scan, Evo ADS) is ahead mainly on discovery (8 against 4, worth 10.0 points at weight 25). **Site ranking: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Discovery (weight 25) | Vetting before use (weight 20) | Runtime control of tool calls (weight 20) | Deployment (weight 15) | Governance and audit (weight 5) | Transparency (weight 10) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [Snyk (Agent Scan, Evo ADS)](/vendors/snyk/) | 77.0 | [8](/vendors/snyk/#evidence-discovery) | [9](/vendors/snyk/#evidence-vetting) | [6](/vendors/snyk/#evidence-runtime) | [8](/vendors/snyk/#evidence-deploy) | [4](/vendors/snyk/#evidence-governance) | [9](/vendors/snyk/#evidence-transparency) | [8](/vendors/snyk/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 71.5 | [7](/vendors/cisco/#evidence-discovery) | [8](/vendors/cisco/#evidence-vetting) | [6](/vendors/cisco/#evidence-runtime) | [8](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-governance) | [7](/vendors/cisco/#evidence-transparency) | [6](/vendors/cisco/#evidence-proof) | | 3rd | [Harmonic Security](/vendors/harmonic-security/) | 69.0 | [8](/vendors/harmonic-security/#evidence-discovery) | [5](/vendors/harmonic-security/#evidence-vetting) | [8](/vendors/harmonic-security/#evidence-runtime) | [8](/vendors/harmonic-security/#evidence-deploy) | [8](/vendors/harmonic-security/#evidence-governance) | [4](/vendors/harmonic-security/#evidence-transparency) | [6](/vendors/harmonic-security/#evidence-proof) | | 4th | [Knostic](/vendors/knostic/) | 61.0 | [6](/vendors/knostic/#evidence-discovery) | [7](/vendors/knostic/#evidence-vetting) | [5](/vendors/knostic/#evidence-runtime) | [7](/vendors/knostic/#evidence-deploy) | [6](/vendors/knostic/#evidence-governance) | [6](/vendors/knostic/#evidence-transparency) | [5](/vendors/knostic/#evidence-proof) | | 5th | [Backslash Security](/vendors/backslash-security/) | 58.5 | [6](/vendors/backslash-security/#evidence-discovery) | [7](/vendors/backslash-security/#evidence-vetting) | [6](/vendors/backslash-security/#evidence-runtime) | [5](/vendors/backslash-security/#evidence-deploy) | [6](/vendors/backslash-security/#evidence-governance) | [4](/vendors/backslash-security/#evidence-transparency) | [6](/vendors/backslash-security/#evidence-proof) | | 6th | [Pillar Security](/vendors/pillar-security/) | 57.0 | [6](/vendors/pillar-security/#evidence-discovery) | [6](/vendors/pillar-security/#evidence-vetting) | [6](/vendors/pillar-security/#evidence-runtime) | [6](/vendors/pillar-security/#evidence-deploy) | [4](/vendors/pillar-security/#evidence-governance) | [4](/vendors/pillar-security/#evidence-transparency) | [6](/vendors/pillar-security/#evidence-proof) | | 7th | [Lasso Security](/vendors/lasso-security/) | 56.0 | [6](/vendors/lasso-security/#evidence-discovery) | [6](/vendors/lasso-security/#evidence-vetting) | [6](/vendors/lasso-security/#evidence-runtime) | [4](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-governance) | [4](/vendors/lasso-security/#evidence-transparency) | [7](/vendors/lasso-security/#evidence-proof) | | 8th | [Microsoft](/vendors/microsoft/) | 54.5 | [8](/vendors/microsoft/#evidence-discovery) | [2](/vendors/microsoft/#evidence-vetting) | [4](/vendors/microsoft/#evidence-runtime) | [6](/vendors/microsoft/#evidence-deploy) | [5](/vendors/microsoft/#evidence-governance) | [8](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 9th | [Noma Security](/vendors/noma-security/) | 51.5 | [6](/vendors/noma-security/#evidence-discovery) | [6](/vendors/noma-security/#evidence-vetting) | [5](/vendors/noma-security/#evidence-runtime) | [4](/vendors/noma-security/#evidence-deploy) | [5](/vendors/noma-security/#evidence-governance) | [3](/vendors/noma-security/#evidence-transparency) | [6](/vendors/noma-security/#evidence-proof) | | 10th | [Bay](/vendors/bay/) | 46.0 | [4](/vendors/bay/#evidence-discovery) | [4](/vendors/bay/#evidence-vetting) | [6](/vendors/bay/#evidence-runtime) | [6](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-governance) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 11th | [Zenity](/vendors/zenity/) | 44.0 | [5](/vendors/zenity/#evidence-discovery) | [3](/vendors/zenity/#evidence-vetting) | [6](/vendors/zenity/#evidence-runtime) | [3](/vendors/zenity/#evidence-deploy) | [5](/vendors/zenity/#evidence-governance) | [3](/vendors/zenity/#evidence-transparency) | [7](/vendors/zenity/#evidence-proof) | | 12th | [CrowdStrike](/vendors/crowdstrike/) | 41.0 | [4](/vendors/crowdstrike/#evidence-discovery) | [3](/vendors/crowdstrike/#evidence-vetting) | [3](/vendors/crowdstrike/#evidence-runtime) | [6](/vendors/crowdstrike/#evidence-deploy) | [5](/vendors/crowdstrike/#evidence-governance) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 13th | [Koi](/vendors/koi/) | 39.0 | [4](/vendors/koi/#evidence-discovery) | [5](/vendors/koi/#evidence-vetting) | [2](/vendors/koi/#evidence-runtime) | [6](/vendors/koi/#evidence-deploy) | [2](/vendors/koi/#evidence-governance) | [1](/vendors/koi/#evidence-transparency) | [8](/vendors/koi/#evidence-proof) | | 14th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 36.5 | [4](/vendors/sentinelone/#evidence-discovery) | [3](/vendors/sentinelone/#evidence-vetting) | [3](/vendors/sentinelone/#evidence-runtime) | [4](/vendors/sentinelone/#evidence-deploy) | [3](/vendors/sentinelone/#evidence-governance) | [4](/vendors/sentinelone/#evidence-transparency) | [6](/vendors/sentinelone/#evidence-proof) | ## Who should pick what With no particular situation, the published ranking applies and Snyk (Agent Scan, Evo ADS) comes first. For a specific situation the weights change, and so can the order: - **[An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/)**: pick Snyk (Agent Scan, Evo ADS) (76.0). Bay is 10th at 49.0. At scale the first problem is knowing which servers exist on which machines, and the tool has to deploy to every laptop. - **[A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/)**: pick Snyk (Agent Scan, Evo ADS) (78.5). Bay is 10th at 48.0. Speed to an inventory rewards discovery breadth and light deployment; control can follow. - **[A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/)**: pick Snyk (Agent Scan, Evo ADS) (77.5). Bay is 9th at 47.5. An approved list depends on assessing each server before use, so vetting weighs most. - **[A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/)**: pick Harmonic Security (72.5). Bay is 8th at 52.5. When a tool call can delete or send data, stopping it on the machine as it happens matters most. - **[A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/)**: pick Cisco (73.0). Bay is 10th at 52.5. Auditors want the record, so the audit log and per-team policy weigh most after discovery. ## What would change this ranking **Top pair.** Cisco rising from 7 to 10 on discovery (weight 25, so each point moves a total by 2.5) would put Cisco above Snyk (Agent Scan, Evo ADS); the gap is 5.5 points. **Bay's next place.** Bay rising from 4 to 7 on discovery (weight 25, so each point moves a total by 2.5) would put Bay above Noma Security; the gap is 5.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which MCP server security tool ranks first? Snyk (Agent Scan, Evo ADS) ranks first with 77.0 of 100 on the site-wide weights. The table above shows the other 13 totals, and each vendor page lists the seven criterion scores with the reason and the facts behind them. A first place here means the best documented coverage of finding servers, vetting them, blocking tool calls and deploying to laptops. It does not mean the product suits every team, and the situation pages show where the order changes. ## Which tools discover and control MCP servers on machines (not gateways)? The 14 scored tools are those whose public documentation describes finding MCP servers on laptops or in developer tools and acting on them there. Some also offer a gateway; only their on-machine capabilities are scored. The weights are Discovery 25, Vetting before use 20, Runtime control of tool calls 20, Deployment 15, Transparency 10, Governance and audit 5 and Proof 5. Discovery and control together carry 65 points because that is the job a platform engineer is hiring the tool for. ## How should you read a score? A vendor's own claim can earn at most 7 on any criterion. A score of 8 to 10 needs a second source, such as third-party reporting or something observed in public documentation, code or a release record. Where evidence was missing the score is 0 and the page says none found on 16 Nov 2026. Breadth figures such as module counts feed only Proof, which weighs 5, so they carry at most 10 percent of any total. ## Related pages - [Method: weights, anchors and rules](/method/) - [Rankings by situation](/best-for/) - [Vendor profiles](/vendors/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) - [Guides for buyers](/guides/) - [Glossary](/glossary/) --- # How MCP security tools are scored URL: https://mcpservercybersecurity.com/method/ Updated: 2026-11-16 Every vendor followed one research plan on 16 Nov 2026, producing 1,469 facts in 32 vendor files, 733 of them in this site's 14. The 1,273 facts with a quote were each checked against their pages. One scorer then scored the 14 vendors blind, and a second reviewer checked the result blind. The weights were committed first. ## The weights The weights were fixed on 2026-11-16, before any score existed, and have not changed since. **Criteria and weights (out of 100)** | Criterion | What it measures | Weight | |---|---|---| | Discovery | Finding the MCP servers configured on machines and in developer tools (desktop AI apps, IDEs, command-line agents), with the servers, tools and clients named, as documented | 25 | | Vetting before use | Assessing a server before it runs: scanning code, tool descriptions and permissions, reputation or provenance, an approved list or registry, as documented | 20 | | Runtime control of tool calls | Allowing, blocking or requiring approval for an MCP tool call on the machine as it happens, with arguments inspected, as documented | 20 | | Deployment | How it deploys on developer machines (agentless, MDM script, endpoint agent, IDE extension, command-line scanner), operating systems, time to deploy | 15 | | Governance and audit | Policy per team or server, an audit log of servers and tool calls, export to SIEM | 5 | | Transparency | Public documentation readable without sign-in, a published price, published research or open-source code | 10 | | Proof | Named customers, funding, independent analyst coverage | 5 | ## Weights by situation **Weights used for each situation view** | Situation | Discovery | Vetting before use | Runtime control of tool calls | Deployment | Governance and audit | Transparency | Proof | |---|---|---|---|---|---|---|---| | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 30 | 15 | 20 | 20 | 5 | 5 | 5 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 40 | 10 | 5 | 30 | 5 | 5 | 5 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 15 | 40 | 15 | 10 | 10 | 5 | 5 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 15 | 15 | 40 | 10 | 10 | 5 | 5 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 25 | 15 | 15 | 5 | 30 | 5 | 5 | ## Score anchors - **0**: none found on the read date - **1-3**: mentioned, partial, or limited to one surface, one layer or one action type - **4-6**: documented and working, with gaps the documentation shows - **7**: the most a vendor claim or vendor documentation alone can earn: detailed, specific, covering the whole criterion - **8-10**: detailed and confirmed by a second source (third-party write-up, or observed directly in public documentation, code, a marketplace listing or a release record) ## Rules - Every vendor is scored 0 to 10 on every criterion, whole numbers, from the evidence file only, by a scorer that works from coded packets with every vendor name removed. - A vendor claim scores at most 7 until a second source confirms it (kind third-party or observed). - None found scores 0 and is written "none found on the read date". - Breadth facts (module counts, integration counts, headcount, funding) inform Proof only, which weighs 5, so they carry at most 10 percent of any ranking. - The total is the weighted sum of the scores divided by 10 (0 to 100 scale, published to one decimal). - Equal totals print as a joint rank (for example joint 2nd); inside a joint rank vendors are listed alphabetically for display only. - Once the first score exists, no weight and no criterion changes. A score changes only when a sourced fact changes, logged with the fact. ## How are MCP security tools scored? Each vendor is scored from 0 to 10 on seven criteria, and the total is the weighted sum divided by 10, shown on a 0 to 100 scale to one decimal. Scores are whole numbers and come from the vendor's evidence file only. The criteria and weights are Discovery 25, Vetting before use 20, Runtime control of tool calls 20, Deployment 15, Transparency 10, Governance and audit 5 and Proof 5. They were written and committed on 16 Nov 2026 before any evidence file or score existed, and none has changed since the first score. ## What is the evidence base? A single research plan was applied to every vendor on 16 Nov 2026. It produced 1,469 facts across 32 vendor files, 733 of them in this site's 14 files. Of the 1,469, 196 are none-found records with no quote, and the other 1,273 were each quote checked against their pages. Each fact stores the claim, the quoted passage, the source URL, the read date and a kind: vendor, third-party, observed or none found. A quote check compared every stored quote with the page it came from. A fact whose quote could not be found on the page was removed or corrected before scoring. ## How does blind scoring work? One scorer read each evidence file with the vendor shown under a letter code rather than a name, and scored it against the anchors. A second reviewer then checked every score, also blind, and any disagreement was settled by the cited facts. The anchors are fixed: 0 for none found, 1 to 3 for a mention or a partial, single-surface capability, 4 to 6 for documented and working with visible gaps, 7 as the most a vendor claim alone can earn, and 8 to 10 when a second source confirms the claim. ## Why are gateways out of scope? Gateways are out of scope because the reader's job here is to find and control the servers installed on machines, and a gateway only sees traffic routed through it. A server started from a config file on a laptop never touches a gateway unless someone forces it to. Where a vendor ships a gateway next to an on-machine component, only the on-machine capability is scored. Ranking gateways is a separate comparison that this site does not publish. ## How are scores corrected? A score changes only when a sourced fact changes, and the change is logged with the fact on the changelog the method page prints. Weights and criteria do not change after the first score. To report an error, write to editor@mcpservercybersecurity.com with the page and the source that contradicts it. The corrections page explains how reports are handled. ## Changelog - 2026-11-16: weights and criteria fixed. - 2026-11-16: evidence read for every vendor by one research plan. - 2026-11-16: scores published (blind: one scorer, then one blind reviewer). - 2026-11-16: this site built from the committed scores. ## Related pages - [The ranking these weights produce](/ranking/) - [Vendor profiles with evidence per score](/vendors/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [Glossary](/glossary/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # MCP security vendors URL: https://mcpservercybersecurity.com/vendors/ Updated: 2026-11-16 Fourteen vendors are profiled here: Backslash Security, Bay, Cisco, CrowdStrike, Harmonic Security, Knostic, Koi, Lasso Security, Microsoft, Noma Security, Pillar Security, SentinelOne, Snyk and Zenity. Each profile lists seven criterion scores, the reason for each and the facts read on 16 Nov 2026. ## All 14 vendors **Vendors in the ranking, with their rank and total** | Vendor | Rank | Total (of 100) | Product assessed | |---|---|---|---| | [Backslash Security](/vendors/backslash-security/) | 5th | 58.5 | Backslash Agentic Endpoint Security platform: MCP Security module (discovery, assessment, governance, endpoint MCP proxy) plus the free MCP Server Security Hub | | [Bay](/vendors/bay/) | 10th | 46.0 | Bay Endpoint inventory and MCP server scanning with Agent Runtime Enforcement | | [Cisco](/vendors/cisco/) | 2nd | 71.5 | Open-source Cisco MCP Scanner (cisco-ai-defense/mcp-scanner) and DefenseClaw skill, plugin and MCP server scanning | | [CrowdStrike](/vendors/crowdstrike/) | 12th | 41.0 | Falcon Guardian endpoint agent and MCP server discovery (AIDR) | | [Harmonic Security](/vendors/harmonic-security/) | 3rd | 69.0 | Harmonic Command with the Local MCP Gateway delivered by the Harmonic Protect Agent (Cloud MCP Gateway in early access) | | [Knostic](/vendors/knostic/) | 4th | 61.0 | Kirin (endpoint client and IDE extension for coding agents and MCP servers) and AgentMesh (scanning catalog for MCP servers, skills and extensions) | | [Koi](/vendors/koi/) | 13th | 39.0 | Cortex Agentic Endpoint Security (Koi), standalone or as a Cortex XDR module | | [Lasso Security](/vendors/lasso-security/) | 7th | 56.0 | Lasso MCP Security (MCP server discovery, risk scoring, tool-call monitoring) and open-source MCP Gateway | | [Microsoft](/vendors/microsoft/) | 8th | 54.5 | Defender for Endpoint local AI agent discovery (MCP server configuration) and AI agent runtime protection, Agent 365 real-time protection for MCP tools | | [Noma Security](/vendors/noma-security/) | 9th | 51.5 | Noma MCP Security (AI-SPM discovery plus Agent Access Control registry plus AI-DR tool poisoning detection) | | [Pillar Security](/vendors/pillar-security/) | 6th | 57.0 | Agentic Endpoint (MCP inventory, risk scores, tool allow-lists) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 14th | 36.5 | Prompt Security Agentic AI Security, MCP discovery and risk scoring (SentinelOne) | | [Snyk (Agent Scan, Evo ADS)](/vendors/snyk/) | 1st | 77.0 | Evo Agentic Development Security (Agent Supply Chain Security, Agent Behavior Governance with MCP Governance) and the open-source Snyk Agent Scan (formerly Invariant Labs mcp-scan); Snyk Code, Open Source and other AppSec engines not counted | | [Zenity](/vendors/zenity/) | 11th | 44.0 | Zenity MCP Security (live view and gateway-based control of MCP servers and tools) plus Zenity coding and personal agent security for Claude Code, Cursor, GitHub Copilot and ChatGPT Codex | ## Related pages - [Ranking of all vendors](/ranking/) - [How vendors are scored](/method/) - [Best vendor by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Backslash Security MCP security: endpoint proxy and vetting hub URL: https://mcpservercybersecurity.com/vendors/backslash-security/ Updated: 2026-11-16 Backslash Security says it discovers every MCP server installed or connected across the endpoint fleet, including personal and unsanctioned ones, with the agents using them and the tools and permissions they expose. Its endpoint MCP proxy inspects traffic between agents and servers and blocks malicious instructions and risky tool calls in real time, per the vendor. On 2025-06-16 Backslash published a free MCP Server Security Hub that scores servers by risk, needs no login and shows GitHub publisher verification. Deployment covers macOS and Windows, and a customer reported rollout in under 30 days. Gaps on 16 Nov 2026: public documentation, a price and a per-client MCP config list are none found, and most capability statements are the vendor's own. ## Strength and weakness for MCP Server Security **Strength:** Endpoint MCP discovery plus an on-machine MCP proxy that blocks risky tool calls, with a public free vetting hub (backslash-security-01, backslash-security-08, backslash-security-06). **Weakness:** No public docs, price or per-client MCP config list; none found on 2026-11-16 (backslash-security-12, backslash-security-16, backslash-security-20). ## Scores and evidence Backslash Security is 5th in the overall ranking at 58.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Backslash Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 6 of 10 | 25 | The vendor claims discovery of all MCP servers, agents, tools and permissions and names several clients; a per-client config list: none found. | [backslash-security-01](#fact-backslash-security-01), [backslash-security-02](#fact-backslash-security-02), [backslash-security-04](#fact-backslash-security-04) | | Vetting before use | 7 of 10 | 20 | The vendor documents assessment for vulnerabilities and permissions, a free scanner with publisher verification and tool-level allowlists; vendor claims only. | [backslash-security-05](#fact-backslash-security-05), [backslash-security-06](#fact-backslash-security-06), [backslash-security-10](#fact-backslash-security-10), [backslash-security-03](#fact-backslash-security-03) | | Runtime control of tool calls | 6 of 10 | 20 | An endpoint MCP proxy is claimed to block risky tool calls in real time; an approval mode and the argument fields inspected: none found. | [backslash-security-08](#fact-backslash-security-08), [backslash-security-09](#fact-backslash-security-09), [backslash-security-11](#fact-backslash-security-11), [backslash-security-12](#fact-backslash-security-12) | | Deployment | 5 of 10 | 15 | A customer reports Mac and Windows rollout in under 30 days; install method and Linux support: none found. | [backslash-security-14](#fact-backslash-security-14), [backslash-security-13](#fact-backslash-security-13), [backslash-security-16](#fact-backslash-security-16) | | Governance and audit | 6 of 10 | 5 | The vendor claims server policies, an audit trail and SIEM integration; per-team scoping: none found. | [backslash-security-17](#fact-backslash-security-17), [backslash-security-18](#fact-backslash-security-18), [backslash-security-19](#fact-backslash-security-19) | | Transparency | 4 of 10 | 10 | Research is published and confirmed by a third party; a price and public documentation: none found. | [backslash-security-21](#fact-backslash-security-21), [backslash-security-22](#fact-backslash-security-22), [backslash-security-06](#fact-backslash-security-06), [backslash-security-20](#fact-backslash-security-20) | | Proof | 6 of 10 | 5 | Third-party press confirms Series A funding and one named customer case study; analyst coverage: none found. | [backslash-security-24](#fact-backslash-security-24), [backslash-security-25](#fact-backslash-security-25), [backslash-security-26](#fact-backslash-security-26) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 5th | 58.5 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 5th | 58.5 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 7th | 57.0 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 5th | 62.0 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 4th | 59.5 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | joint 5th | 60.0 | ## Product, deployment, price and owner - Product assessed: Backslash Agentic Endpoint Security platform: MCP Security module (discovery, assessment, governance, endpoint MCP proxy) plus the free MCP Server Security Hub - Deployment: Endpoint-based platform with an endpoint MCP proxy; the exposure assessment is agentless and read-only - Time to deploy: Customer Happy Returns reported under 30 days for a full Mac and Windows deployment - Price: none found on 2026-11-16 - Owner: Backslash Security (standalone; no parent named on its pages) - Website: [backslash.security](https://www.backslash.security) ## Facts and sources - **backslash-security-01** (vendor): Backslash says it discovers every MCP server installed or connected across the endpoint fleet, including personal and unsanctioned ones, the agents using them, and the tools and permissions they expose.Quote: "Discover every MCP server installed or connected across your endpoints fleet, including personal and unsanctioned MCPs, the agents using them, and the tools and permissions they expose."Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-02** (vendor): Backslash says it discovers public, private, locally developed, remote and internally hosted MCP servers on employee endpoints.Quote: "Backslash discovers public, private, locally developed, remote, and internally hosted MCP servers running across employee endpoints."Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-03** (vendor): Backslash's December 2025 press release lists centralized MCP discovery on developer workstations, by AI agents and in IDEs, plus vetting for vulnerabilities, malware, configuration and permission risks.Quote: "Centralized discovery of MCPs in use on developer workstations, by AI agents, and in IDEs"Source: [backslash.security](https://www.backslash.security/press-releases/backslash-security-introduces-comprehensive-mcp-security-solution-for-vibe-coding-and-ai-native-software-development), read 2026-11-16. - **backslash-security-04** (vendor): Backslash names Claude, GitHub Copilot, Antigravity, Cursor and OpenClaw as tools it gives visibility across, and Claude Code, Cursor, Codex and Devin appear in its console example; it publishes no per-client MCP config list.Quote: "Get full visibility and risk posture assessment, across all tools including Claude, GitHub Copilot, Antigravity, Cursor and OpenClaw."Source: [backslash.security](https://www.backslash.security/agentic-endpoint-security), read 2026-11-16. - **backslash-security-05** (vendor): Backslash says it continuously assesses MCP servers for vulnerabilities, malicious behavior, excessive permissions, configuration risks and supply chain issues.Quote: "Continuously assess MCP servers for vulnerabilities, malicious behavior, excessive permissions, configuration risks, and supply chain issues"Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-06** (vendor): Backslash published a free MCP Server Security Hub on 2025-06-16 that scores MCP servers by risk, requires no login, and shows attack vector and GitHub publisher verification.Quote: "The hub (for short) is free to use and does not require logging in."Source: [backslash.security](https://www.backslash.security/blog/mcp-server-security-hub), read 2026-11-16. - **backslash-security-07** (vendor): Backslash says it has assessed more than 65,000 MCP servers.Quote: "65,000+ MCP Servers Assessed (and Counting). Is Yours Safe?"Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-08** (vendor): Backslash says its endpoint MCP proxy inspects communication between agents and MCP servers in real time and blocks malicious instructions, risky tool calls and policy violations.Quote: "Backslash's endpoint MCP proxy inspects communication between AI agents and MCP servers in real time, blocking malicious instructions, risky tool calls, excessive permissions, and policy violations."Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-09** (vendor): Backslash says its MCP Proxy intercepts inbound and outbound MCP activity in real time to block data leakage and prompt injection attempts.Quote: "MCP Proxy that intercepts both inbound and outbound activities in real time, to block data leakage and prompt injection attempts"Source: [backslash.security](https://www.backslash.security/press-releases/backslash-security-introduces-comprehensive-mcp-security-solution-for-vibe-coding-and-ai-native-software-development), read 2026-11-16. - **backslash-security-10** (vendor): Backslash says it can vet, allowlist and monitor MCPs at the tool level and block unsafe ones.Quote: "Block, vet, allowlist and monitor MCPs at tool level."Source: [backslash.security](https://www.backslash.security/agentic-endpoint-security), read 2026-11-16. - **backslash-security-11** (vendor): Backslash's homepage describes risky agent actions on developer workstations being blocked inline before execution, with example blocked actions including credential access and data sent to unapproved destinations.Quote: "Risky agent actions blocked inline, before execution"Source: [backslash.security](https://www.backslash.security), read 2026-11-16. - **backslash-security-12** (none-found): None found on 2026-11-16: no published description of a per-call approval prompt (ask the user) mode, or of the tool-call arguments fields inspected, on the MCP security page or the endpoint security page.Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-13** (vendor): Backslash says its MCP security needs zero developer configuration.Quote: "Zero-configuration deployment that requires no developer effort"Source: [backslash.security](https://www.backslash.security/press-releases/backslash-security-introduces-comprehensive-mcp-security-solution-for-vibe-coding-and-ai-native-software-development), read 2026-11-16. - **backslash-security-14** (vendor): Customer Happy Returns, a UPS company, says it deployed Backslash across all Mac and Windows endpoints and was running in under 30 days.Quote: "We did a full-scale deployment across all our Mac and Windows endpoints and were fully running in under 30 days."Source: [backslash.security](https://www.backslash.security/case-studies/happy-returns), read 2026-11-16. - **backslash-security-15** (vendor): Backslash's endpoint exposure assessment is described as agentless and read-only.Quote: "Agentless Read-only No data retention"Source: [backslash.security](https://www.backslash.security/ai-endpoint-exposure-assessment), read 2026-11-16. - **backslash-security-16** (none-found): None found on 2026-11-16: no public install guide, MDM or script method, or Linux support statement for the endpoint product on its site; docs are not public.Source: [backslash.security](https://www.backslash.security/agentic-endpoint-security), read 2026-11-16. - **backslash-security-17** (vendor): Backslash says it enforces policies governing which MCP servers can be used, how they are configured and what access they have across employee endpoints.Quote: "Define and enforce security policies that govern which MCP servers can be used, how they're configured, and what access they're allowed across employee endpoints."Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-18** (vendor): Backslash says it keeps a forensic audit trail of agent prompts, processes and actions.Quote: "Capture full tracing of agent prompts, processes and actions, giving security teams a complete audit trail to support forensic investigation, incident response, and compliance."Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-19** (vendor): Backslash says audit of events and policy violations is integrated into SIEM and SOC tools.Quote: "Ongoing audit of events and policy violations for compliance and forensics, integrated into SIEM and SOC tools"Source: [backslash.security](https://www.backslash.security/press-releases/backslash-security-introduces-comprehensive-mcp-security-solution-for-vibe-coding-and-ai-native-software-development), read 2026-11-16. - **backslash-security-20** (none-found): None found on 2026-11-16: no public pricing page (the /pricing URL returns 404) and no public product documentation on backslash.security; only a Book a Demo path.Quote: "Page Not Found"Source: [backslash.security](https://www.backslash.security/pricing), read 2026-11-16. - **backslash-security-21** (vendor): Backslash published research on public MCP servers (press release 2025-06-25) and a free security hub, and lists SOC 2 Type II attestation.Quote: "SOC 2 Type II Compliant & Attested"Source: [backslash.security](https://www.backslash.security/use-cases/mcp-security), read 2026-11-16. - **backslash-security-22** (third-party): The Register reported on 2025-07-21 that Backslash researchers found ways around Cursor's command denylist in auto-run mode, showing Backslash publishes agent security research.Quote: "Cursor’s denylist cannot be relied upon"Source: [theregister.com](https://www.theregister.com/2025/07/21/cursor_ai_safeguards_easily_bypassed/), read 2026-11-16. - **backslash-security-23** (vendor): On 2026-02-10 Backslash announced a $19 million Series A led by KOMPAS VC, following an $8 million seed.Quote: "today announced a $19 million Series A funding round led by KOMPAS VC"Source: [backslash.security](https://www.backslash.security/press-releases/backslash-security-raises-19m-series-a-to-secure-vibe-coding-boom-in-the-enterprise-bolsters-board-with-cybersecurity-industry-leader), read 2026-11-16. - **backslash-security-24** (third-party): SiliconANGLE reports Backslash Security Inc. was founded in 2022 and that its $19M Series A was led by KOMPAS VC.Quote: "Founded in 2022, Backslash focuses on protecting what it describes as the emerging vibe coding attack surface."Source: [siliconangle.com](https://siliconangle.com/2026/02/10/backslash-security-lands-19m-tackle-security-risks-emerging-vibe-coding/), read 2026-11-16. - **backslash-security-25** (vendor): Happy Returns, a UPS company, is a named customer with a published case study and quote.Quote: "Backslash gave us a live picture of the AI our engineers were already running, and a way to govern it without slowing anyone down"Source: [backslash.security](https://www.backslash.security/), read 2026-11-16. - **backslash-security-26** (none-found): None found on 2026-11-16: no independent analyst report on Backslash (Gartner, Forrester, IDC) read; its site lists a Latio resource and a Gartner Hype Cycle mention in news titles only, not read.Source: [backslash.security](https://www.backslash.security/customers), read 2026-11-16. - **backslash-security-27** (none-found): None found on 2026-11-16: no Backslash press release or news item dated 2026-08-18 or later on its news page; the latest listed is 2026-08-02 (Agentic Fabric Graph, announced 2026-07-29).Source: [backslash.security](https://www.backslash.security/news), read 2026-11-16. - **backslash-security-28** (vendor): Backslash describes itself as the Agentic AI Endpoint Security platform for AI agents, MCP servers and Skills on employee endpoints; no parent company is named on its pages.Quote: "Backslash Security is the Agentic AI Endpoint Security platform."Source: [backslash.security](https://www.backslash.security/blog/mcp-server-security-hub), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Zenity profile](/vendors/zenity/) - [Bay profile](/vendors/bay/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Bay MCP security: endpoint scan and on-device enforcement URL: https://mcpservercybersecurity.com/vendors/bay/ Updated: 2026-11-16 Bay says its endpoint product finds agents, extensions, CLI tools and MCP servers on employee machines through one script run by an existing MDM or EDR, and that every MCP server is scanned for malicious indicators before approval and each time it changes. Third-party press reported that Bay Enforcer, a static binary, decides allow, ask or deny on each action on the device through a native hook, with an install under five minutes. Bay says policy follows identity-provider groups, decisions and transcripts are logged, and events reach a SIEM in OCSF format. Bay is a young company that left stealth in 2026, publishes no price and has little independent testing behind its claims. A list of MCP clients read and inspection of tool-call arguments: none found on 16 Nov 2026. Bay is a young company that left stealth in 2026, publishes no price and has little independent testing behind its claims. ## Strength and weakness for MCP Server Security **Strength:** Scans MCP servers for malicious indicators before approval and enforces tool-call decisions on the device (bay-26, bay-48). **Weakness:** None found on 2026-11-16 for a list of MCP clients discovered or tool-call argument inspection (bay-71, bay-57). ## Scores and evidence Bay is 10th in the overall ranking at 46.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Bay: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 4 of 10 | 25 | The vendor claims discovery of MCP servers and names Claude Code and Cursor; a list of MCP client configurations read: none found. | [bay-14](#fact-bay-14), [bay-15](#fact-bay-15), [bay-70](#fact-bay-70), [bay-71](#fact-bay-71) | | Vetting before use | 4 of 10 | 20 | The vendor claims scanning before approval for malicious indicators; detail and an enforced server registry: none found. | [bay-26](#fact-bay-26), [bay-27](#fact-bay-27), [bay-77](#fact-bay-77) | | Runtime control of tool calls | 6 of 10 | 20 | Third-party press confirms local allow, ask or deny on each action via a native hook; inspection of tool-call arguments: none found. | [bay-46](#fact-bay-46), [bay-47](#fact-bay-47), [bay-54](#fact-bay-54), [bay-55](#fact-bay-55), [bay-57](#fact-bay-57) | | Deployment | 6 of 10 | 15 | A single script through MDM or EDR, with an under-five-minute install reported by a third party; supported OS list: none found. | [bay-16](#fact-bay-16), [bay-17](#fact-bay-17), [bay-19](#fact-bay-19), [bay-21](#fact-bay-21) | | Governance and audit | 7 of 10 | 5 | The vendor documents IdP-group policy, logged decisions, full transcripts and OCSF SIEM export, all vendor-stated. | [bay-30](#fact-bay-30), [bay-31](#fact-bay-31), [bay-61](#fact-bay-61), [bay-62](#fact-bay-62), [bay-76](#fact-bay-76) | | Transparency | 0 of 10 | 10 | Public documentation, a published price and published research or open-source code: none found. | [bay-34](#fact-bay-34), [bay-35](#fact-bay-35), [bay-36](#fact-bay-36) | | Proof | 7 of 10 | 5 | Third-party press confirms $50 million in funding and names customers; analyst coverage: none found. | [bay-06](#fact-bay-06), [bay-07](#fact-bay-07), [bay-08](#fact-bay-08), [bay-09](#fact-bay-09) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 10th | 46.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 10th | 49.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 10th | 48.0 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 9th | 47.5 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 8th | 52.5 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 10th | 52.5 | ## Product, deployment, price and owner - Product assessed: Bay Endpoint inventory and MCP server scanning with Agent Runtime Enforcement - Deployment: Agentless: one script through an existing MDM or EDR runs Wave, an ephemeral binary; Bay Enforcer, a static binary, decides on the device via native hook; no proxy (vendor claims) - Time to deploy: Under 5 minutes, as claimed by the vendor; proof of value in under 14 days (SecurityWeek quote) - Price: none found on 2026-11-16 - Owner: Bay AI, Inc. - Website: [bay.io](https://bay.io/?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor) ## Facts and sources - **bay-01** (vendor): Bay AI, Inc. publishes the Bay site and states its copyright on every page.Quote: "© 2026 Bay AI, Inc. All rights reserved."Source: [bay.io](https://www.bay.io/about?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-02** (vendor): Bay says it was founded by three engineers from endpoint and API security.Quote: "Bay was founded by three engineers who spent their careers solving the hardest problems in endpoint and API security."Source: [bay.io](https://www.bay.io/about?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-03** (third-party): On 2026-11-10 Bay named its leadership: Yuval Zacharia is CEO and co-founder, Aner Morag is CPO and Boris Danilovich is VP of R&D, per SiliconANGLE.Quote: "Zacharia founded Bay with Chief Product Officer Aner Morag and Vice President of Research and Development Boris Danilovich."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-04** (vendor): Bay says it was founded in 2026 and is headquartered in Tel Aviv; its press release is dated 2026-11-10.Quote: "Founded in 2026, Bay is headquartered in Tel Aviv and serves customers in the United States and Europe."Source: [bay.io](https://www.bay.io/blog/bay-emerges-from-stealth-with-50m-after-fortune-100-companies-used-its-platform-to-triple-autonomous-ai-agent-use-2?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-05** (vendor): On 2026-11-10 Bay said it had raised $50 million in total funding.Quote: "Bay has raised $50 million in total funding."Source: [bay.io](https://www.bay.io/blog/bay-emerges-from-stealth-with-50m-after-fortune-100-companies-used-its-platform-to-triple-autonomous-ai-agent-use-2?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-06** (third-party): On 2026-11-11 SecurityWeek reported Bay emerged from stealth with a $37 million Series A led by Battery Ventures plus a $13 million seed.Quote: "emerged from stealth with $50 million funding (comprising a $37 million Series A round led by Battery Ventures"Source: [securityweek.com](https://www.securityweek.com/bay-raises-50-million-to-control-what-enterprise-ai-agents-can-do/), read 2026-11-16. - **bay-07** (third-party): On 2026-11-10 SiliconANGLE reported Micron, Wix, ClickUp and Peloton are among dozens of Bay enterprise customers.Quote: "Micron Technology Inc., Wix.com Ltd., ClickUp Inc. and Peloton Interactive Inc. are among the dozens of enterprise customers signed in the nine months since Bay was founded."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-08** (third-party): The Next Web noted that Bay did not name its Fortune 100 customers or say how it measured the tripling claim, and that one named customer is also an investor.Quote: "The company did not name those customers or say how it measured the figure, and TNW has not verified it."Source: [thenextweb.com](https://thenextweb.com/news/bay-50m-ai-agent-security-stealth-battery-ventures), read 2026-11-16. - **bay-09** (none-found): None found on 2026-11-16: independent analyst coverage of Bay (Gartner, Forrester or similar) on the Bay home page or blog.Source: [bay.io](https://bay.io/?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-10** (vendor): Bay says it is SOC 2 and ISO 27001 certified.Quote: "SOC 2 & ISO 27001 Certified"Source: [bay.io](https://www.bay.io/about?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-14** (vendor): Bay says its endpoint product finds agents, extensions, CLI tools, MCP servers, skills and models on employee devices.Quote: "Bay finds every agent, extension, CLI tool, MCP server, skill and model on employee devices."Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-15** (vendor): Bay says its first scan maps agents, MCP servers, extensions, models and skills on devices, including unapproved ones.Quote: "On the first scan, Bay maps every agent, MCP server, extension, model and skill on devices, including ones nobody approved."Source: [bay.io](https://www.bay.io/platform?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-16** (vendor): Bay says it deploys with one script through an existing MDM, with no proxy and nothing installed by users.Quote: "Bay deploys with one script through an existing MDM. It does not require a proxy to route traffic through and does not require anything to be installed by users."Source: [bay.io](https://www.bay.io/platform?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-17** (vendor): Bay says it is agentless and deploys via MDM or EDR using Wave, an ephemeral binary, with no persistent agent and no proxy.Quote: "Bay is agentless: it deploys through your existing MDM or EDR using Wave, an ephemeral binary, so there’s no persistent agent to maintain and no proxy in the path."Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-18** (vendor): Bay says the endpoint inventory appears in under 5 minutes.Quote: "Bay brings it all into view in under 5 minutes and keeps it in bounds, without slowing anyone down."Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-19** (third-party): On 2026-11-10 SiliconANGLE reported Bay installs through the customer's existing MDM or EDR and the company says it takes under five minutes.Quote: "Bay installs through the mobile device management or endpoint detection and response tools a customer already has, and the company says it takes under five minutes."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-20** (third-party): On 2026-11-11 SecurityWeek quoted Bay saying a fleet deploys in under five minutes and a proof of value completes in under 14 days.Quote: "complete a proof of value in less than 14 days."Source: [securityweek.com](https://www.securityweek.com/bay-raises-50-million-to-control-what-enterprise-ai-agents-can-do/), read 2026-11-16. - **bay-21** (none-found): None found on 2026-11-16: a list of supported operating systems for the Bay endpoint scanner, on the endpoint product page and the platform page.Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-25** (vendor): Bay lists an approval workflow for newly found AI tools; details of the workflow were not published on the page.Quote: "An approval workflow people actually use"Source: [bay.io](https://www.bay.io/shadow-ai-discovery?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-26** (vendor): Bay says every plugin, package, skill and MCP server is scanned and vetted before approval and each time it changes.Quote: "Every plugin, package, skill and MCP server is scanned and vetted before it’s approved and every time it changes."Source: [bay.io](https://www.bay.io/shadow-ai-discovery?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-27** (vendor): Bay says every agent, package, extension, MCP server and skill is scanned for malicious indicators before approval.Quote: "Every agent, package, extension, MCP server and skill is scanned for malicious indicators before its approved."Source: [bay.io](https://www.bay.io/platform?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-30** (vendor): Bay says it connects to the identity provider so policies apply to the right people and groups.Quote: "Bay connects to your identity provider, so every policy applies to the right people and groups, and keeps up as your organization changes."Source: [bay.io](https://www.bay.io/ai-governance-compliance?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-31** (vendor): Bay says policies can be scoped by IdP group, device or runtime and can set which MCP servers may read or write.Quote: "Policies can be scoped by IdP group, device or runtime, and can set which MCP servers can read or write"Source: [bay.io](https://www.bay.io/platform?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-34** (observed): Bay documentation at docs.bay.io redirects to a sign-in page and is open only to signed-in Bay users.Quote: "The Bay documentation is open only to signed-in Bay users."Source: [docs.bay.io](https://docs.bay.io/?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-35** (none-found): None found on 2026-11-16: a published price. The site has no pricing page (bay.io/pricing returns 404) and the demo page only offers a 30-minute demo.Source: [bay.io](https://www.bay.io/demo?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-36** (none-found): None found on 2026-11-16: published research, threat reports or open-source code; the Bay blog lists only two announcement posts dated 2026-11-10.Source: [bay.io](https://www.bay.io/blog?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-38** (vendor): Bay says teams run Claude Code, Codex, Gemini CLI and GitHub Copilot in cloud workspaces and that Bay covers them.Quote: "Teams now run agents like Claude Code, Codex, Gemini CLI and GitHub Copilot in clusters, cloud workspaces and hosted sessions"Source: [bay.io](https://www.bay.io/platform/cloud-workspaces?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-46** (third-party): On 2026-11-10 SiliconANGLE reported the AI tool hands each proposed action to Bay Enforcer, a static binary, through a native hook on the machine.Quote: "Each time an agent proposes an action, the AI tool running it hands the request to Bay Enforcer, a static binary, through a native hook."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-47** (vendor): Bay says runtime enforcement decides allow, ask, deny or steer on every action, locally and in milliseconds with no cloud round-trip.Quote: "Bay reads each agent’s whole session to understand its intent, then decides on every action: allow, ask, deny or steer."Source: [bay.io](https://www.bay.io/agent-runtime-enforcement?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-48** (vendor): Bay says it acts on credential access, risky shell commands and untrusted MCP server connections in milliseconds.Quote: "When an agent reaches for credentials, runs a risky shell command or connects to an untrusted MCP server, Bay allows, asks, denies or steers the action in milliseconds"Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-54** (third-party): On 2026-11-11 SecurityWeek reported Bay decides in milliseconds to allow, refer to the user, redirect or deny each action.Quote: "decides in milliseconds whether to allow the action, refer it to the user, redirect it based on policy, or simply deny it"Source: [securityweek.com](https://www.securityweek.com/bay-raises-50-million-to-control-what-enterprise-ai-agents-can-do/), read 2026-11-16. - **bay-55** (third-party): On 2026-11-10 The Next Web reported Bay checks every action an agent proposes before it runs, then allows, blocks or asks a person.Quote: "It checks every action an agent proposes before it runs, then allows it, blocks it or asks a person."Source: [thenextweb.com](https://thenextweb.com/news/bay-50m-ai-agent-security-stealth-battery-ventures), read 2026-11-16. - **bay-57** (none-found): None found on 2026-11-16: named blocking of package installs or network calls, or inspection of tool-call arguments, on the endpoint and runtime pages.Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-58** (vendor): Bay says it applies intent-aware enforcement on every tool call in cloud agents.Quote: "Intent-aware enforcement on every tool call."Source: [bay.io](https://www.bay.io/platform/cloud-workspaces?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-61** (vendor): Bay says every decision is logged against the policy behind it for auditors.Quote: "Every decision is logged against the policy behind it, ready for your auditors."Source: [bay.io](https://www.bay.io/platform?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-62** (vendor): Bay says events reach the customer SIEM in OCSF format.Quote: "Events arrive in your SIEM in OCSF format, with enough context that analysts can triage without opening another console."Source: [bay.io](https://www.bay.io/ai-agent-investigation?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-63** (vendor): Bay says it records every prompt, tool call and system action on one timeline per session.Quote: "Every prompt, tool call and system action on one timeline, so you can see how an agent got from its first instruction to its last action."Source: [bay.io](https://www.bay.io/ai-agent-investigation?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-68** (vendor): Bay's product walkthrough says a discovered Claude Code install was connected to Salesforce, Atlassian and Sentry and could read plaintext secrets in its MCP configs.Quote: "the agent can read plaintext secrets in its MCP configs, as well as AWS and SSH keys on disk"Source: [bay.io](https://www.bay.io/platform?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-69** (vendor): Bay's product walkthrough names a Claude Code install as a discovered finding.Quote: "A Claude Code install is signed in with a personal Gmail account and connected to Salesforce, Atlassian and Sentry."Source: [bay.io](https://www.bay.io/platform?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-70** (vendor): Bay names Claude Code and Cursor as coding agents it secures on the endpoint.Quote: "How do you secure AI coding agents like Claude Code and Cursor on the endpoint?"Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-71** (none-found): None found on 2026-11-16: a list of MCP client configurations discovered (Claude Desktop, VS Code, Windsurf, JetBrains) or of remote server support, on the endpoint page.Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-76** (vendor): Bay says its records include full transcripts of prompts, actions and MCP calls traceable to a person.Quote: "Full transcripts of prompts, actions and MCP calls - traceable to the person behind each session."Source: [bay.io](https://www.bay.io/platform/cloud-workspaces?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-77** (none-found): None found on 2026-11-16: a per-server allowlist or registry for MCP servers beyond scoping which servers may read or write; see approval workflow mention.Source: [bay.io](https://www.bay.io/platform?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Backslash Security profile](/vendors/backslash-security/) - [Cisco profile](/vendors/cisco/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Cisco MCP security: MCP Scanner and AI Defense URL: https://mcpservercybersecurity.com/vendors/cisco/ Updated: 2026-11-16 Cisco's MCP Scanner is Apache-2.0 open-source code that combines the AI Defense inspect API, YARA rules and LLM-based analysis. Its known-configs mode reads the well-known client config locations on a machine, including Windsurf, Cursor, Claude and VS Code, then scans each server's tools, prompts, resources and instructions. Cisco's hooks apply allow, ask or block decisions to tool calls, with user or group policy, recorded decisions and OpenTelemetry export documented; installs run from the command line or an MDM package on Linux, macOS and Windows. Gaps on 16 Nov 2026: a fleet inventory, an approved-server registry on the machine, per-call inspection of MCP tool arguments, and a published price for paid AI Defense are none found. ## Strength and weakness for MCP Server Security **Strength:** A free Apache-2.0 scanner that reads known client configs and scans server tools, prompts, resources, packages and code before use (cisco-21, cisco-22, cisco-24, cisco-30). **Weakness:** None found on 2026-11-16 for per-call MCP argument inspection or an approved-server registry on the machine (cisco-36, cisco-31). ## Scores and evidence Cisco is 2nd in the overall ranking at 71.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Cisco: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 7 of 10 | 25 | The open-source scanner reads configs for Windsurf, Cursor, Claude and VS Code and scans the tools found; a fleet inventory and Claude Code configs: none found. | [cisco-24](#fact-cisco-24), [cisco-25](#fact-cisco-25), [cisco-33](#fact-cisco-33), [cisco-30](#fact-cisco-30) | | Vetting before use | 8 of 10 | 20 | The vendor scans tools, prompts and packages with YARA and LLM analysis at install, and third-party press confirms scanning and server blocking. | [cisco-21](#fact-cisco-21), [cisco-22](#fact-cisco-22), [cisco-23](#fact-cisco-23), [cisco-34](#fact-cisco-34), [cisco-31](#fact-cisco-31), [cisco-33](#fact-cisco-33) | | Runtime control of tool calls | 6 of 10 | 20 | Hooks check tool calls with allow, ask or block, and third-party press confirms server blocking; inspection of MCP tool-call arguments specifically: none found. | [cisco-10](#fact-cisco-10), [cisco-12](#fact-cisco-12), [cisco-31](#fact-cisco-31), [cisco-32](#fact-cisco-32), [cisco-36](#fact-cisco-36) | | Deployment | 8 of 10 | 15 | Command-line and MDM-package installs on Linux, macOS and Windows, with a five-minute install reported by a third party. | [cisco-11](#fact-cisco-11), [cisco-27](#fact-cisco-27), [cisco-38](#fact-cisco-38), [cisco-33](#fact-cisco-33) | | Governance and audit | 8 of 10 | 5 | User or group policy, recorded decisions and OpenTelemetry export are documented, and a third party reports a SIEM connector. | [cisco-16](#fact-cisco-16), [cisco-17](#fact-cisco-17), [cisco-18](#fact-cisco-18), [cisco-35](#fact-cisco-35), [cisco-37](#fact-cisco-37) | | Transparency | 7 of 10 | 10 | Public documentation and observed Apache-2.0 open-source code; a paid price: none found. | [cisco-29](#fact-cisco-29), [cisco-30](#fact-cisco-30), [cisco-56](#fact-cisco-56) | | Proof | 6 of 10 | 5 | A public company with third-party launch coverage and one named customer voice; analyst coverage for this product: none found. | [cisco-60](#fact-cisco-60), [cisco-58](#fact-cisco-58), [cisco-59](#fact-cisco-59) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 2nd | 71.5 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 3rd | 71.5 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 3rd | 73.5 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 2nd | 74.0 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 3rd | 69.0 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 1st | 73.0 | ## Product, deployment, price and owner - Product assessed: Open-source Cisco MCP Scanner (cisco-ai-defense/mcp-scanner) and DefenseClaw skill, plugin and MCP server scanning - Deployment: Network-level enforcement without agents or libraries for AI Defense; Kong plugin, Python SDK, command-line MCP Scanner, and DefenseClaw hooks installed per user or by MDM package - Time to deploy: Cisco says DefenseClaw installs in about five minutes (SiliconANGLE); none found on 2026-11-16 for AI Defense - Price: none found on 2026-11-16 (paid AI Defense); Explorer Edition is described as free - Owner: Cisco Systems, Inc. - Website: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html) ## Facts and sources - **cisco-01** (vendor): Cisco Systems, Inc. is the owner and copyright holder of DefenseClaw, the open-source agent governance layer published in the cisco-ai-defense GitHub organization alongside AI Defense.The exact quote is on the source page.Source: [github.com](https://github.com/cisco-ai-defense/defenseclaw), read 2026-11-16. - **cisco-04** (vendor): Cisco says AI Defense enforces at the network level, with no agents or libraries needed on the application side.Quote: "Cisco enforces AI security at the network-level without the need for agents or libraries."Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-10** (vendor): DefenseClaw documentation says its guardrail checks a prompt, a shell command or a tool call before the agent acts and answers allow, alert, ask or block.Quote: "The guardrail checks what your agent is about to do (a prompt, a shell command, a tool call) against rules, and answers allow, alert, ask you, or block."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/guardrail/), read 2026-11-16. - **cisco-11** (vendor): DefenseClaw documentation says support on Linux, macOS and Windows is marked Supported only after a live test verified hooks and a block before the tool call ran.Quote: "means a live test on that OS verified the agent's hooks and a block before the tool call ran."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/support-matrix/), read 2026-11-16. - **cisco-12** (vendor): The DefenseClaw compatibility table lists hook contracts for Claude Code that inspect prompts, tool calls and tool results, with similar rows for Codex, Cursor, GitHub Copilot CLI and others.Quote: "Claude Code | `>=2.1.154, <2.1.219` | `claudecode-hooks-v1` / `v7` | 28 | prompt, tool_call, tool_result, event_content"Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/connectors/compatibility/), read 2026-11-16. - **cisco-16** (vendor): On 2026-10-21, Cisco said enterprise administrators can set user- or group-based policy for which coding agents run on managed endpoints and which run only in a sandbox.Quote: "Enterprise administrators can configure user or group-based policy to determine which coding agents can run on their managed endpoints"Source: [blogs.cisco.com](https://blogs.cisco.com/ai/protect-enterprise-coding-agents-with-cisco-defenseclaw-and-nvidia-openshell), read 2026-11-16. - **cisco-17** (vendor): On 2026-10-21, Cisco said every block, approval and decision is recorded with the agent, sandbox and policy behind it and can be sent to Splunk or any OpenTelemetry backend.Quote: "Every block, approval and decision is recorded along with the agent, the sandbox and the policy behind it, and can be sent to Splunk or any OpenTelemetry backend."Source: [blogs.cisco.com](https://blogs.cisco.com/ai/protect-enterprise-coding-agents-with-cisco-defenseclaw-and-nvidia-openshell), read 2026-11-16. - **cisco-18** (vendor): DefenseClaw observability documentation lists OTLP, Splunk, Galileo, file and webhook destinations for logs, traces and metrics.Quote: "goes to the destinations you add: an OpenTelemetry (OTLP) collector, Splunk, Galileo, a file, or a webhook."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/observability/), read 2026-11-16. - **cisco-21** (vendor): Cisco's open-source MCP Scanner combines the AI Defense inspect API, YARA rules and LLM-based analysis to detect malicious MCP tools.Quote: "The MCP Scanner combines Cisco AI Defense inspect API, YARA rules and LLM-based analysis to detect malicious MCP tools."Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-22** (vendor): The MCP Scanner scans a server's tools, prompts, resources and instructions.Quote: "Scan MCP tools, prompts, resources, and server instructions for security findings"Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-23** (vendor): The MCP Scanner can scan PyPI packages for hidden behavior inside a Docker sandbox before they are used.Quote: "Download and scan PyPI packages inside a Docker sandbox with behavioral analysis."Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-24** (vendor): The MCP Scanner has a known-configs mode that scans servers from well-known MCP client config locations on the machine.Quote: "scan servers from well-known client config locations on this machine"Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-25** (vendor): The MCP Scanner README names Windsurf, Cursor, Claude and VS Code as the clients whose config paths it reads.Quote: "Scan well-known MCP config paths (Windsurf, Cursor, Claude, VS Code)"Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-26** (vendor): The MCP Scanner can scan a remote MCP server over SSE or streamable HTTP when given its URL.Quote: "remote**: scan a remote MCP server (SSE or streamable HTTP)."Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-27** (vendor): The MCP Scanner runs as a stand-alone command-line tool or a REST API server, installed with uv.Quote: "Run scanner as a stand-alone CLI tool or REST API server"Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-28** (vendor): The MCP Scanner README documents a static mode for CI/CD pipelines and air-gapped environments.Quote: "Scan pre-generated JSON files without live server connections - perfect for CI/CD pipelines and air-gapped environments"Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-29** (vendor): DefenseClaw documentation says its public website is the source of truth for install, configuration and operator workflows, and the code is Apache-2.0.Quote: "is the source of truth for installation, setup, configuration, commands, and operator workflows"Source: [github.com](https://github.com/cisco-ai-defense/defenseclaw), read 2026-11-16. - **cisco-30** (observed): The GitHub API lists the cisco-ai-defense/mcp-scanner repository as Apache-2.0 licensed with 1119 stars and a push on 2026-11-15.Quote: "mcp-scanner |pushed 2026-11-15 |stars 1119 |license Apache-2.0"Source: [api.github.com](https://api.github.com/orgs/cisco-ai-defense/repos), read 2026-11-16. - **cisco-31** (third-party): On 2026-03-23, SiliconANGLE reported DefenseClaw lets administrators block specific MCP servers, with block rules applied in two seconds without restarting agents.Quote: "A permission management feature enables administrators to block specific MCP servers."Source: [siliconangle.com](https://siliconangle.com/2026/03/23/cisco-debuts-new-ai-agent-security-features-open-source-defenseclaw-tool/), read 2026-11-16. - **cisco-32** (third-party): SiliconANGLE quoted Cisco saying that blocking an MCP server removes its endpoint from the sandbox network allow-list so OpenShell denies all connections.Quote: "When you block an MCP server, the endpoint is removed from the sandbox network allow-list and OpenShell denies all connections."Source: [siliconangle.com](https://siliconangle.com/2026/03/23/cisco-debuts-new-ai-agent-security-features-open-source-defenseclaw-tool/), read 2026-11-16. - **cisco-33** (third-party): SiliconANGLE reported developers can install DefenseClaw in about five minutes and that it scans MCP tools and plug-ins an agent uses.Quote: "Cisco says developers can install DefenseClaw in about five minutes."Source: [siliconangle.com](https://siliconangle.com/2026/03/23/cisco-debuts-new-ai-agent-security-features-open-source-defenseclaw-tool/), read 2026-11-16. - **cisco-34** (vendor): Cisco says DefenseClaw scans skills, plugins and MCP servers at install, before they are allowed into the environment.Quote: "it gets scanned before it is allowed into your environment."Source: [blogs.cisco.com](https://blogs.cisco.com/ai/defenseclaw-is-live), read 2026-11-16. - **cisco-35** (vendor): Cisco says DefenseClaw streams every scan result, block decision, tool call and alert as structured events.The exact quote is on the source page.Source: [blogs.cisco.com](https://blogs.cisco.com/ai/defenseclaw-is-live), read 2026-11-16. - **cisco-36** (none-found): None found on 2026-11-16: a statement that DefenseClaw or the MCP Scanner inspects the arguments of an individual MCP tool call on the machine and blocks it; the docs name tool_call hooks for coding agents in general, not MCP tool calls specifically.Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/connectors/compatibility/), read 2026-11-16. - **cisco-37** (third-party): SiliconANGLE reported DefenseClaw sends telemetry on the risks it finds to Splunk through a prepackaged connector.Quote: "DefenseClaw sends telemetry about the cybersecurity risks that it finds to Splunk through a prepackaged connector."Source: [siliconangle.com](https://siliconangle.com/2026/03/23/cisco-debuts-new-ai-agent-security-features-open-source-defenseclaw-tool/), read 2026-11-16. - **cisco-38** (vendor): DefenseClaw documentation describes an enterprise edition that an administrator deploys with a package or an MDM and that users cannot turn off.Quote: "the managed standalone profile an administrator deploys with a package or an MDM. Users can't turn it off."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/support-matrix/), read 2026-11-16. - **cisco-41** (vendor): Cisco says its planned MCP gateway applies authorization policies across tools and blocks out-of-bounds agent actions, but also says the features are in development.Quote: "the MCP gateway applies authorization policies consistently across tools and services."Source: [blogs.cisco.com](https://blogs.cisco.com/security/security-agentic-ai-how-cisco-brings-zero-trust-to-your-new-digital-workforce), read 2026-11-16. - **cisco-56** (none-found): None found on 2026-11-16: a published price for Cisco AI Defense on its product page; the free Explorer Edition is announced but no paid list price was found.Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-57** (third-party): CRN reported that at RSAC 2026 Cisco introduced a free Explorer Edition of AI Defense for red teaming agents.Quote: "Cisco also debuted updates to its AI Defense offering with the introduction of a free “Explorer Edition.”"Source: [crn.com](https://www.crn.com/news/security/2026/cisco-unveils-zero-trust-for-ai-agents-5-things-to-know), read 2026-11-16. - **cisco-58** (vendor): The AI Defense product page quotes Kent Noyes of World Wide Technology, a named customer voice.Quote: "The adoption of AI exposes companies to new risks that traditional cybersecurity solutions don't address."Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-59** (none-found): None found on 2026-11-16: independent analyst coverage (Gartner, Forrester, IDC) or a funding round for AI Defense on the pages read; Cisco is a public company.Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-60** (third-party): SiliconANGLE covered the 2026-03-23 DefenseClaw release at RSAC 2026 and the Explorer Edition.Quote: "The networking giant also released an open-source tool, DefenseClaw, that can scan AI agents for vulnerabilities."Source: [siliconangle.com](https://siliconangle.com/2026/03/23/cisco-debuts-new-ai-agent-security-features-open-source-defenseclaw-tool/), read 2026-11-16. ## News mentioning Cisco - [Cisco DefenseClaw 1.0.0 adds managed installs and inventory](/news/cisco-defenseclaw-1-0-0-managed-enterprise-release/) ## Related pages - [All vendor profiles](/vendors/) - [Bay profile](/vendors/bay/) - [CrowdStrike profile](/vendors/crowdstrike/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # CrowdStrike MCP security: Falcon Guardian URL: https://mcpservercybersecurity.com/vendors/crowdstrike/ Updated: 2026-11-16 CrowdStrike says Falcon Guardian discovers shadow AI, agents, MCP servers, skills and tools on endpoints through the Falcon sensor on Windows, macOS and Linux, and finds malicious AI skills and MCP servers before they enter the AI environment. It says Guardian links prompts, skill use, tool calls, MCP servers and identity to downstream execution for supported agents, and agent hooks block malicious activity. Agent-type allow rules and SIEM export are documented, and Falcon AI Gateway is included with Guardian at no additional cost. Gaps on 16 Nov 2026: named MCP client discovery, an MCP server allowlist, per-call argument inspection, a Guardian price and a time to deploy are none found; a third party says MCP vulnerability scanning is planned. ## Strength and weakness for MCP Server Security **Strength:** Falcon Guardian discovers MCP servers and malicious MCP servers and skills on endpoints, and links MCP use to execution (facts 1, 15, 41). **Weakness:** None found on 2026-11-16 for named client discovery, an MCP allowlist, or per-call argument inspection (facts 58, 59, 61). ## Scores and evidence CrowdStrike is 12th in the overall ranking at 41.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **CrowdStrike: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 4 of 10 | 25 | The vendor claims endpoint discovery of MCP servers and tools on Windows, macOS and Linux; named MCP client configurations: none found. | [crowdstrike-01](#fact-crowdstrike-01), [crowdstrike-02](#fact-crowdstrike-02), [crowdstrike-41](#fact-crowdstrike-41), [crowdstrike-59](#fact-crowdstrike-59) | | Vetting before use | 3 of 10 | 20 | The vendor claims finding malicious MCP servers, with MCP vulnerability scanning only planned per a third party; an enforced registry: none found. | [crowdstrike-15](#fact-crowdstrike-15), [crowdstrike-07](#fact-crowdstrike-07), [crowdstrike-61](#fact-crowdstrike-61) | | Runtime control of tool calls | 3 of 10 | 20 | Agent hooks block malicious activity generally; per-MCP-tool-call control with arguments inspected: none found. | [crowdstrike-10](#fact-crowdstrike-10), [crowdstrike-11](#fact-crowdstrike-11), [crowdstrike-58](#fact-crowdstrike-58) | | Deployment | 6 of 10 | 15 | An endpoint sensor via device management on Windows, macOS and Linux; a time to deploy: none found. | [crowdstrike-02](#fact-crowdstrike-02), [crowdstrike-30](#fact-crowdstrike-30), [crowdstrike-60](#fact-crowdstrike-60) | | Governance and audit | 5 of 10 | 5 | Agent-type allow rules and SIEM export are documented; a per-team MCP server policy: none found. | [crowdstrike-12](#fact-crowdstrike-12), [crowdstrike-38](#fact-crowdstrike-38), [crowdstrike-39](#fact-crowdstrike-39), [crowdstrike-61](#fact-crowdstrike-61) | | Transparency | 4 of 10 | 10 | Endpoint bundle prices, research and observed open-source MCP servers are published; a price for this product and public documentation: none found. | [crowdstrike-45](#fact-crowdstrike-45), [crowdstrike-48](#fact-crowdstrike-48), [crowdstrike-50](#fact-crowdstrike-50), [crowdstrike-46](#fact-crowdstrike-46), [crowdstrike-47](#fact-crowdstrike-47) | | Proof | 7 of 10 | 5 | A large vendor with third-party acquisition coverage and a customer story; analyst coverage here is in another market and vendor-cited. | [crowdstrike-05](#fact-crowdstrike-05), [crowdstrike-43](#fact-crowdstrike-43), [crowdstrike-44](#fact-crowdstrike-44) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 12th | 41.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 12th | 42.5 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 11th | 46.5 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 13th | 39.0 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 12th | 39.0 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 12th | 42.5 | ## Product, deployment, price and owner - Product assessed: Falcon Guardian endpoint agent and MCP server discovery (AIDR) - Deployment: Falcon sensor on endpoints with agent-level hooks, a routed Falcon AI Gateway for apps and workloads beyond the sensor, and device-management rollout for the Cowork integration - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 for Falcon Guardian; endpoint bundles are priced (Falcon Go $7.99 per device monthly) and Falcon AI Gateway is included with Guardian at no additional cost - Owner: CrowdStrike Holdings, Inc. (public company; CrowdStrike sells the products itself) - Website: crowdstrike.com ## Facts and sources - **crowdstrike-01** (vendor): CrowdStrike says Falcon Guardian discovers shadow AI, agents, applications, models, MCP servers, skills and tools on endpoints.Quote: "Discover shadow AI, agents, applications, models, MCP servers, skills, and tools."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-02** (vendor): On 2026-09-01 CrowdStrike introduced Falcon Guardian, the successor to Falcon AI Detection and Response (AIDR), and says it discovers AI agents on supported Windows, macOS and Linux endpoints.Quote: "Falcon Guardian continuously discovers known and previously unknown AI agents across supported Windows, macOS, and Linux endpoints."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-03** (vendor): CrowdStrike names Falcon Guardian as the evolution of Falcon AI Detection and Response (AIDR) and its flagship AI security product.Quote: "the evolution of Falcon AI Detection and Response (AIDR) and CrowdStrike’s flagship solution for the AIDR category"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-04** (vendor): CrowdStrike lists George Kurtz as CEO and Founder and Michael Sentonas as President on its executive team page.Quote: "George Kurtz CEO & Founder Michael Sentonas President"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/about-us/executive-team/), read 2026-11-16. - **crowdstrike-05** (third-party): SiliconANGLE reported on 2026-11-03 that CrowdStrike Holdings plans to buy Fiddler AI and Dreadnode, two AI agent security startups, announced with a Falcon Guardian update.Quote: "CrowdStrike Holdings Inc. today disclosed plans to buy Fiddler AI Inc. and Dreadnode Inc., two startups that specialize in securing artificial intelligence agents."Source: [siliconangle.com](https://siliconangle.com/2026/11/03/crowdstrike-to-acquire-cybersecurity-startups-fiddler-ai-dreadnode-for-reported-495m/), read 2026-11-16. - **crowdstrike-06** (vendor): On 2026-11-03 CrowdStrike announced its intent to acquire Fiddler, to become Falcon Observe for AI agent observability, generally available in Q4 FY27 (a future product, not scored as available).Quote: "Fiddler will become part of the Falcon platform as Falcon Observe, generally available in Q4 FY27."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-to-acquire-fiddler/), read 2026-11-16. - **crowdstrike-07** (third-party): SiliconANGLE says Dreadnode, being acquired by CrowdStrike and planned for integration into Falcon Guardian, looks for vulnerabilities in the MCP servers an agent uses (planned, not yet part of Guardian).Quote: "Dreadnode also looks for vulnerabilities in the MCP servers that an AI agent uses to carry out its work."Source: [siliconangle.com](https://siliconangle.com/2026/11/03/crowdstrike-to-acquire-cybersecurity-startups-fiddler-ai-dreadnode-for-reported-495m/), read 2026-11-16. - **crowdstrike-10** (vendor): On 2026-11-03 CrowdStrike said agent-level hooks for Claude Code, OpenAI Codex, GitHub Copilot, OpenCode and Cursor inspect prompts and block malicious activity before harmful actions run (hooks act on the agent, not specifically on MCP tool calls).Quote: "Hooks for Claude Code, OpenAI Codex, GitHub Copilot, OpenCode, and Cursor enable organizations to inspect prompts and block malicious activity at the source, before harmful actions are executed."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-11** (vendor): On 2026-09-01 CrowdStrike said Falcon Guardian can drive automatic containment by blocking malicious agent behaviors and compromised assets at runtime.Quote: "drive automatic containment by blocking malicious agent behaviors and compromised assets at runtime across agents"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-12** (vendor): On 2026-09-01 CrowdStrike said Falcon Guardian lets organizations define which supported agent types may run on managed endpoints and blocks unauthorized types.Quote: "helping security teams sanction approved agents and prevent unauthorized agent types from running."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-15** (vendor): On 2026-11-03 CrowdStrike said Falcon Guardian discovers malicious AI skills and MCP servers to identify threats before they enter the AI environment.Quote: "Falcon Guardian expands AI supply chain protection by discovering malicious AI skills and MCP servers to help security teams identify threats before they enter the AI environment."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-21** (vendor): CrowdStrike says developers connect an existing SDK or coding tool to the gateway by changing the base URL and using a gateway key.Quote: "Developers connect an existing SDK or coding tool by changing the base URL to point at the gateway and using a gateway key."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/), read 2026-11-16. - **crowdstrike-23** (vendor): On 2026-11-03 CrowdStrike said Falcon AI Gateway is included with Falcon Guardian at no additional cost.Quote: "Included with Falcon Guardian at no additional cost, it extends AI security to applications and workloads beyond direct endpoint sensor coverage, including unmanaged environments."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-introduces-ai-kill-switch/), read 2026-11-16. - **crowdstrike-27** (vendor): On 2026-11-03 CrowdStrike described pre-beta coverage of cloud workload and Kubernetes agents, including LangChain and LangGraph apps.Quote: "In cloud workloads and Kubernetes, Falcon Guardian will extend these capabilities to enterprise-developed AI agents and applications, including those built with LangChain and LangGraph."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-30** (vendor): On 2026-11-03 CrowdStrike said admins turn on the Cowork integration through existing device management, as they deploy the Falcon sensor.Quote: "Admins will turn the integration on through their existing device management, the same way they deploy the Falcon sensor today"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-secures-ai-agents-inside-claude-cowork/), read 2026-11-16. - **crowdstrike-38** (vendor): On 2026-09-01 CrowdStrike said Guardian exports agent telemetry into Falcon Next-Gen SIEM as first-party data.Quote: "Falcon Guardian natively exports agent telemetry into CrowdStrike Falcon® Next-Gen SIEM as first-party data, pre-mapped to its schema for immediate correlation, detection, and automation"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-39** (vendor): On 2026-11-03 CrowdStrike listed Guardian additions: Falcon Data Replicator export of prompt activity, Flight Control support, and token consumption dashboards.Quote: "Falcon Data Replicator integration for prompt activity, Falcon Flight Control support, new executive summary and token consumption dashboards, and expanded Windows AI application coverage."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-41** (vendor): CrowdStrike says Guardian links prompts, skill use, tool calls, MCP servers and identity to downstream execution for supported agents.Quote: "connect prompts, skill use, tool calls, MCP servers, and identity with downstream system execution for supported agents."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-43** (vendor): CrowdStrike publishes a customer story saying Cogentrix gained day-one AI visibility with Falcon Guardian (customer claim, no time figure).Quote: "See how Cogentrix gains day-one AI visibility and governs employee and agentic AI with Falcon Guardian."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-44** (vendor): CrowdStrike lists on its blog that it was named a Leader in the 2026 IDC MarketScape for Worldwide Augmented Identity Security Solutions (dated 2026-11-12; identity security, not AI security specifically).Quote: "CrowdStrike Named a Leader in the 2026 IDC MarketScape for Worldwide Augmented Identity Security Solutions"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-introduces-ai-kill-switch/), read 2026-11-16. - **crowdstrike-45** (vendor): CrowdStrike publishes list prices for endpoint bundles, for example Falcon Go at $7.99 per device billed monthly; Guardian is not on that page.Quote: "Falcon Go $7.99 per device, billed monthly"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/pricing/), read 2026-11-16. - **crowdstrike-46** (none-found): None found on 2026-11-16: a published price for Falcon Guardian or Falcon AIDR on the pricing page (only endpoint bundles are priced; other products say contact us).Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/pricing/), read 2026-11-16. - **crowdstrike-47** (none-found): None found on 2026-11-16: public product documentation for Falcon Guardian; the documentation address returned a Falcon login page.Source: [falcon.crowdstrike.com](https://falcon.crowdstrike.com/documentation/), read 2026-11-16. - **crowdstrike-48** (vendor): On 2026-01-09 CrowdStrike published research explaining AI tool poisoning through MCP tool descriptions and how to defend against it.Quote: "In this blog, we explain how AI tool poisoning works, the different forms it can take, and how organizations can strengthen their defenses against this type of attack."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/), read 2026-11-16. - **crowdstrike-50** (observed): CrowdStrike publishes falcon-mcp, an open-source MIT-licensed MCP server for Falcon; it is a server that agents use, not an MCP security scanner.Quote: "is a Model Context Protocol (MCP) server that connects AI agents with"Source: [github.com](https://github.com/CrowdStrike/falcon-mcp), read 2026-11-16. - **crowdstrike-51** (observed): CrowdStrike publishes aidr-mcp-server on GitHub, which integrates with CrowdStrike AIDR APIs using an API token.Quote: "server that provides integration with CrowdStrike AIDR APIs."Source: [github.com](https://github.com/CrowdStrike/aidr-mcp-server), read 2026-11-16. - **crowdstrike-53** (vendor): CrowdStrike sells a Shadow AI Visibility Service (expert-led, launched 2026-04-21) that discovers AI tools, agents, copilots and extensions across endpoint, cloud and SaaS.Quote: "Discover AI tools, agents, copilots, extensions, and model-connected services operating across endpoint, cloud, and SaaS"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/services/ai-security-services/shadow-ai-visibility/), read 2026-11-16. - **crowdstrike-58** (none-found): None found on 2026-11-16: a documented control that allows, blocks or asks approval for a single MCP tool call with its arguments inspected, on the Falcon Guardian page.Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-59** (none-found): None found on 2026-11-16: named discovery of MCP servers configured in Claude Desktop, Cursor, VS Code or Claude Code, or of remote MCP servers, on the Falcon Guardian page.Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-60** (none-found): None found on 2026-11-16: a published time to deploy for Falcon Guardian on the Guardian page.Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-61** (none-found): None found on 2026-11-16: an approved MCP server list or registry enforced per team, on the Falcon Guardian page.Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-63** (third-party): TechRadar reported that CrowdStrike confirmed it would acquire identity security startup SGNL for about $740 million, to support continuous access control for human, non-human and AI identities (article undated in extracted text).Quote: "CrowdStrike has confirmed it will be acquiring identity security startup SGNL for around $740 million in a mostly-cash deal, subject to regulatory approval."Source: [techradar.com](https://www.techradar.com/pro/crowdstrike-acquires-continuous-identity-firm-sgnl-for-usd740m-in-push-for-next-gen-security), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Cisco profile](/vendors/cisco/) - [Harmonic Security profile](/vendors/harmonic-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Harmonic Security MCP security: local gateway control URL: https://mcpservercybersecurity.com/vendors/harmonic-security/ Updated: 2026-11-16 Harmonic Security's endpoint agent installs a local MCP gateway on the machine. Its documentation shows the agent inventorying clients (Claude Code, Codex, Cursor, Windsurf, VS Code and Copilot CLI) and servers, denying new servers until they are approved, and enforcing policy on each tool call with blocking and access requests. Harmonic says policies can be set per agent, server, tool and action, and its docs describe access groups per server by group or department and full invocation logs with parameters and responses. Installers cover Windows, macOS and Linux and are pushed by Intune, Jamf or Kandji. Gaps on 16 Nov 2026: pre-use scanning of server code or tool descriptions, a published price and Claude Desktop discovery are none found, and the gateway setting is off by default. ## Strength and weakness for MCP Server Security **Strength:** Local MCP gateway that inventories clients and servers, inspects tool calls and blocks per agent, server, tool and action with identity-aware access groups and invocation logs (harmonic-security-48, harmonic-security-54, harmonic-security-55). **Weakness:** None found on 2026-11-16 for pre-use scanning of MCP server code or tool descriptions, and the gateway setting is off by default and the Cloud gateway is early access (harmonic-security-46, harmonic-security-52). ## Scores and evidence Harmonic Security is 3rd in the overall ranking at 69.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Harmonic Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 8 of 10 | 25 | The endpoint agent inventories MCP clients (Claude Code, Codex, Cursor, Windsurf, VS Code, Copilot CLI) and servers, and observed docs show clients, servers and tool calls in use. Claude Desktop is not named. | [harmonic-security-12](#fact-harmonic-security-12), [harmonic-security-14](#fact-harmonic-security-14), [harmonic-security-41](#fact-harmonic-security-41), [harmonic-security-43](#fact-harmonic-security-43), [harmonic-security-59](#fact-harmonic-security-59) | | Vetting before use | 5 of 10 | 20 | Observed docs show new servers denied until approved and allow or deny of tools, prompts and resources; pre-use scanning of code, tool descriptions or reputation: none found. | [harmonic-security-44](#fact-harmonic-security-44), [harmonic-security-45](#fact-harmonic-security-45), [harmonic-security-46](#fact-harmonic-security-46) | | Runtime control of tool calls | 8 of 10 | 20 | Observed docs show local enforcement of tool calls with blocking and access requests, and the vendor says each tool call's content is checked against per-tool policy. | [harmonic-security-32](#fact-harmonic-security-32), [harmonic-security-50](#fact-harmonic-security-50), [harmonic-security-48](#fact-harmonic-security-48), [harmonic-security-59](#fact-harmonic-security-59), [harmonic-security-62](#fact-harmonic-security-62) | | Deployment | 8 of 10 | 15 | Observed docs show the endpoint agent installing the local gateway with installers for Windows, macOS and Linux, pushed by Intune, JAMF or Kandji in minutes per the vendor. | [harmonic-security-25](#fact-harmonic-security-25), [harmonic-security-26](#fact-harmonic-security-26), [harmonic-security-51](#fact-harmonic-security-51), [harmonic-security-52](#fact-harmonic-security-52) | | Governance and audit | 8 of 10 | 5 | Observed docs show access groups per server by group or department and full invocation logs with parameters; SIEM is by webhook, a vendor claim. | [harmonic-security-54](#fact-harmonic-security-54), [harmonic-security-55](#fact-harmonic-security-55), [harmonic-security-56](#fact-harmonic-security-56), [harmonic-security-35](#fact-harmonic-security-35) | | Transparency | 4 of 10 | 10 | Documentation is public without sign-in; a published price and open-source code or research: none found. | [harmonic-security-37](#fact-harmonic-security-37), [harmonic-security-38](#fact-harmonic-security-38), [harmonic-security-40](#fact-harmonic-security-40) | | Proof | 6 of 10 | 5 | Third-party reports confirm $76 million in funding and over 100 customers, with one named testimonial; independent analyst coverage: none found. | [harmonic-security-03](#fact-harmonic-security-03), [harmonic-security-04](#fact-harmonic-security-04), [harmonic-security-06](#fact-harmonic-security-06), [harmonic-security-09](#fact-harmonic-security-09) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 3rd | 69.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 2nd | 72.5 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 2nd | 74.0 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 3rd | 65.0 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 1st | 72.5 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 2nd | 72.5 | ## Product, deployment, price and owner - Product assessed: Harmonic Command with the Local MCP Gateway delivered by the Harmonic Protect Agent (Cloud MCP Gateway in early access) - Deployment: Browser extension plus Harmonic Protect Agent (endpoint agent) pushed by MDM; local MCP Gateway installed by the agent; OpenTelemetry and API connectors; no network proxy - Time to deploy: Harmonic says first insight in about 30 minutes and deployment in minutes via Intune, JAMF or Kandji - Price: none found on 2026-11-16: pricing page lists Explore, Guide and Command tiers without amounts - Owner: Harmonic Security (standalone; no parent company stated on its pages, which show 2026 Harmonic Security copyright) - Website: harmonic.security ## Facts and sources - **harmonic-security-01** (vendor): Harmonic Security describes itself as selling AI Governance and Control, a control layer for AI use by employees.Quote: "Harmonic Security delivers AI Governance and Control (AIGC), the intelligent control layer that secures and enables the AI-First workforce."Source: [harmonic.security](https://www.harmonic.security/about), read 2026-11-16. - **harmonic-security-02** (vendor): The about page names Alastair Paterson as CEO and co-founder and Bryan Woolgar-O'Neil as CTO and co-founder, and lists no parent company.Quote: "Alastair Paterson CEO & CO-FOUNDER, HARMONIC Bryan Woolgar-O'Neil CTO & CO-FOUNDER, HARMONIC"Source: [harmonic.security](https://www.harmonic.security/about), read 2026-11-16. - **harmonic-security-03** (third-party): On 2026-10-14, SecurityWeek reported a $50 million raise for Harmonic led by Crosspoint Capital Partners, with the founders previously at Digital Shadows.Quote: "The fresh investment round was led by Crosspoint Capital Partners, with additional support from existing investors"Source: [securityweek.com](https://www.securityweek.com/harmonic-security-raises-50-million-for-workforce-ai-protection/), read 2026-11-16. - **harmonic-security-04** (third-party): SecurityBrief reported that Harmonic's Series B brings total funding to $76 million and that more than 100 companies use the platform.Quote: "Harmonic Security has raised USD $50 million in a Series B funding round led by Crosspoint Capital Partners, bringing its total funding to USD $76 million."Source: [securitybrief.co.uk](https://securitybrief.co.uk/story/harmonic-security-raises-usd-50-million-in-series-b), read 2026-11-16. - **harmonic-security-05** (third-party): On 2026-10-14, Unite.AI reported Harmonic's statement that it protects millions of employees at more than 100 companies.Quote: "Harmonic said it protects and enables AI use for millions of employees at more than 100 companies"Source: [unite.ai](https://www.unite.ai/harmonic-security-raises-50m-series-b-to-expand-us-and-uk-teams/), read 2026-11-16. - **harmonic-security-06** (vendor): Harmonic's about page carries a named customer testimonial from Monolithic Power Systems, an IT security director saying it finds unauthorized AI apps.Quote: "Harmonic Security uncovers blindspots, including unauthorized and risky AI applications my employees are using."Source: [harmonic.security](https://www.harmonic.security/about), read 2026-11-16. - **harmonic-security-08** (third-party): SecurityBrief reported that Harmonic is listed among ecosystem integrations for Cisco Cloud Control's Agent Builder through the Model Context Protocol.Quote: "It is listed among ecosystem integrations for Cisco Cloud Control's Agent Builder through Model Context Protocol."Source: [securitybrief.co.uk](https://securitybrief.co.uk/story/harmonic-security-raises-usd-50-million-in-series-b), read 2026-11-16. - **harmonic-security-09** (none-found): None found on 2026-11-16: independent analyst coverage of Harmonic itself (a Gartner or Forrester page naming Harmonic). The Gartner MCP Gateway page on harmonic.security is a title-only vendor repost and does not describe Harmonic.Source: [harmonic.security](https://www.harmonic.security/resources/gartner-innovation-insight-mcp-gateway), read 2026-11-16. - **harmonic-security-12** (vendor): Harmonic's endpoint page says its Endpoint Agent finds desktop apps, coding assistants, CLI tools, local models and the MCP servers running on developer machines.Quote: "Harmonic covers the full set, from Claude Desktop and ChatGPT Desktop to Cursor, Claude Code, GitHub Copilot, Ollama and the MCP servers running on developer machines."Source: [harmonic.security](https://www.harmonic.security/solutions/endpoint-ai-security), read 2026-11-16. - **harmonic-security-14** (vendor): Harmonic says it automatically inventories MCP clients (Claude Code, Claude Cowork, Codex, Cursor) and the servers they reach.Quote: "Harmonic automatically discovers and inventories MCP clients such as Claude Code, Claude Cowork, Codex and Cursor, along with the servers they reach"Source: [harmonic.security](https://www.harmonic.security/solutions/ai-agent-security), read 2026-11-16. - **harmonic-security-24** (vendor): Harmonic says most teams see their first insight in about 30 minutes.Quote: "Most teams see first insight in about 30 minutes."Source: [harmonic.security](https://www.harmonic.security/solutions/shadow-ai-detection), read 2026-11-16. - **harmonic-security-25** (vendor): Harmonic says deployment takes minutes through Intune, JAMF or Kandji.Quote: "Deployment in minutes via Intune, JAMF, or Kandji."Source: [harmonic.security](https://www.harmonic.security/solutions/shadow-ai-detection), read 2026-11-16. - **harmonic-security-26** (observed): Harmonic's docs list supported operating systems for the Protect Agent.Quote: "Windows 11 macOS 15 (Sequoia), 26 (Tahoe) and 27 (Golden Gate) Ubuntu 22LTS, 24LTS and 26LTS Rocky Linux 9 and 10"Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/protect-agent/installation-and-deployment/frequently-asked-questions), read 2026-11-16. - **harmonic-security-32** (observed): Harmonic's MCP access-control docs say a user's attempt to use a server they lack permission for is blocked and creates an access request for an administrator.Quote: "The attempt is blocked by the Harmonic MCP Gateway"Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/portal-guides/managing-your-organization/configure-mcp-server-access-controls), read 2026-11-16. - **harmonic-security-35** (vendor): Harmonic's pricing page lists a full prompt audit log that can be streamed to a data lake (Command and Guide tiers show it as an add-on).Quote: "Full prompt audit log Stream all interaction data to your data lake for custom analysis"Source: [harmonic.security](https://www.harmonic.security/pricing), read 2026-11-16. - **harmonic-security-37** (observed): Harmonic's product documentation is public without sign-in at docs.harmonicsecurity.app.Quote: "Browse by topic or search across all Harmonic docs."Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/), read 2026-11-16. - **harmonic-security-38** (none-found): None found on 2026-11-16: a published price. The pricing page lists Explore, Guide and Command tiers with features but no amounts.Source: [harmonic.security](https://www.harmonic.security/pricing), read 2026-11-16. - **harmonic-security-40** (none-found): None found on 2026-11-16: open-source code from Harmonic. The GitHub search API for repositories named harmonic-security under the harmonicsecurity org returned zero results.Source: [api.github.com](https://api.github.com/search/repositories?q=harmonic-security+in:name+org:harmonicsecurity), read 2026-11-16. - **harmonic-security-41** (observed): Harmonic's MCP Gateway docs say it shows which MCP clients and servers are in use and logs every tool call.Quote: "Which MCP clients and servers are in use across your organization"Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/mcp-gateway/getting-started/what-is-mcp-gateway), read 2026-11-16. - **harmonic-security-42** (vendor): Harmonic's pricing page lists MCP client and server discovery in the Command tier.Quote: "MCP client and server discovery"Source: [harmonic.security](https://www.harmonic.security/pricing), read 2026-11-16. - **harmonic-security-43** (vendor): Harmonic's CISO guide (2025-11-20) names Cursor, Windsurf and Claude Code as local MCP clients its local gateway covers.Quote: "A Local MCP Gateway governs activity when the MCP Client runs directly on an employee’s machine. such as with tools like Cursor, Windsurf, or Claude Code."Source: [harmonic.security](https://www.harmonic.security/resources/a-cisos-guide-to-mcp-risks), read 2026-11-16. - **harmonic-security-44** (observed): Harmonic's docs say new MCP servers can default to denied until approved, and access groups apply to newly discovered servers.Quote: "You can modify the template to enforce stricter defaults, such as denying access by default so that new servers require explicit approval."Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/portal-guides/managing-your-organization/configure-mcp-server-access-controls), read 2026-11-16. - **harmonic-security-45** (observed): Harmonic's docs let admins allow or deny a server's Tools, Prompts or Resources as categories.Quote: "Allow or deny access to a server's Tools, Prompts, or Resources as categories"Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/portal-guides/managing-your-organization/configure-mcp-server-access-controls), read 2026-11-16. - **harmonic-security-46** (none-found): None found on 2026-11-16: pre-use scanning of MCP server code, tool descriptions or reputation. Looked on the MCP Gateway docs overview, the AI agent security page and the access controls doc.Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/mcp-gateway/getting-started/what-is-mcp-gateway), read 2026-11-16. - **harmonic-security-47** (vendor): Harmonic says calls carrying sensitive data or unallowed actions are stopped and the agent is told why.Quote: "Calls that carry sensitive data or attempt an action you have not allowed are stopped, and the agent is told why."Source: [harmonic.security](https://www.harmonic.security/solutions/ai-agent-security), read 2026-11-16. - **harmonic-security-48** (vendor): Harmonic says policies can be set per agent, server, tool and action.Quote: "Policies can be set per agent, server, tool and action."Source: [harmonic.security](https://www.harmonic.security/solutions/ai-agent-security), read 2026-11-16. - **harmonic-security-49** (vendor): Harmonic's pricing page says Command lets teams define which tools an agent can call and block destructive actions.Quote: "Define exactly which tools an agent can call and block destructive or unauthorized actions"Source: [harmonic.security](https://www.harmonic.security/pricing), read 2026-11-16. - **harmonic-security-50** (vendor): Harmonic says it reads the content of each tool call and checks it against policy.Quote: "Harmonic reads the content of each tool call, checks it against your policy, and records the outcome."Source: [harmonic.security](https://www.harmonic.security/solutions/ai-agent-security), read 2026-11-16. - **harmonic-security-51** (observed): Harmonic's docs say the local MCP Gateway is deployed by installing the Protect Agent, not the gateway directly.Quote: "Deployment of the MCP Gateway is done by installing the Harmonic Protect Agent rather than direct deployment of the gateway itself."Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/mcp-gateway/local-mcp-gateway/installation-and-deployment), read 2026-11-16. - **harmonic-security-52** (observed): Harmonic's docs say gateway deployment is a Protect Agent setting that is off by default, with installers for Windows, Linux and macOS.Quote: "MCP Gateway deployment is controlled by a Protect Agent configuration setting, which is off by default."Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/mcp-gateway/local-mcp-gateway/installation-and-deployment), read 2026-11-16. - **harmonic-security-53** (vendor): Harmonic's October 15, 2025 launch release describes MCP Gateway as a locally installed gateway for developers.Quote: "a developer-friendly, locally installed gateway that gives security teams complete visibility and control over their organization's agentic AI ecosystem"Source: [harmonic.security](https://www.harmonic.security/resources/harmonic-security-launches-mcp-gateway-providing-complete-visibility-and-control-over-an-organizations-agentic-ai-ecosystem), read 2026-11-16. - **harmonic-security-54** (observed): Harmonic's docs describe Access Groups per MCP server, based on employee, group, department or office.Quote: "Access Groups consist of identity conditions and permission settings."Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/portal-guides/managing-your-organization/configure-mcp-server-access-controls), read 2026-11-16. - **harmonic-security-55** (observed): Harmonic's docs say the gateway keeps full invocation logs of every tool call with parameters and responses.Quote: "Full invocation logs showing every tool call, with parameters and responses"Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/mcp-gateway/getting-started/what-is-mcp-gateway), read 2026-11-16. - **harmonic-security-56** (vendor): Harmonic's pricing page lists alert webhooks for SIEM and SOAR integrations in Guide.Quote: "Alert webhooks for SIEM and SOAR integrations"Source: [harmonic.security](https://www.harmonic.security/pricing), read 2026-11-16. - **harmonic-security-57** (vendor): Harmonic's agent security page says invocation logging captures an audit record of every interaction.Quote: "Invocation logging captures a detailed audit record of every interaction for forensics and compliance."Source: [harmonic.security](https://www.harmonic.security/solutions/ai-agent-security), read 2026-11-16. - **harmonic-security-58** (vendor): On 2025-10-15 Harmonic announced its MCP Gateway as a product launch.Quote: "Harmonic Security today announces MCP Gateway"Source: [harmonic.security](https://www.harmonic.security/resources/harmonic-security-launches-mcp-gateway-providing-complete-visibility-and-control-over-an-organizations-agentic-ai-ecosystem), read 2026-11-16. - **harmonic-security-59** (observed): Harmonic's docs say coding-agent inline controls cover prompts and tool calls for Claude Code, Codex, Cursor, VS Code, GitHub Copilot CLI and others.Quote: "Coding agent inline controls covers user prompts and tool calls."Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/portal-guides/detection-and-response/sensitive-data-controls-for-coding-agents-and-desktop-applications), read 2026-11-16. - **harmonic-security-60** (observed): On 2026-08-19 Harmonic released a Cloud MCP Gateway in early access, covering cloud-to-cloud MCP connections with sensitive data inspection on every tool call and result.Quote: "Identity-aware access controls and sensitive data inspection on every tool call and result"Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/release-notes), read 2026-11-16. - **harmonic-security-61** (observed): Harmonic's docs say tool call results are checked for indirect prompt injection (early access).Quote: "Prompt Injection in tool results is known as Indirect Prompt Injection"Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/portal-guides/detection-and-response/prompt-injection-detection-early-access), read 2026-11-16. - **harmonic-security-62** (observed): Harmonic's local gateway docs say policies are enforced locally without a network proxy or cloud round-trip.The exact quote is on the source page.Source: [docs.harmonicsecurity.app](https://docs.harmonicsecurity.app/mcp-gateway/getting-started/what-is-mcp-gateway), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [CrowdStrike profile](/vendors/crowdstrike/) - [Knostic profile](/vendors/knostic/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Knostic MCP security: Kirin and AgentMesh URL: https://mcpservercybersecurity.com/vendors/knostic/ Updated: 2026-11-16 Knostic pairs two pieces. AgentMesh assesses an MCP server before use, scanning for prompt injection and supply-chain threats and returning a verdict with findings. Kirin, an endpoint client deployed by MDM script or direct install with an IDE extension for Cursor and VS Code, inventories servers with publisher and transport and blocks the connection to a server judged bad or unapproved before it is established. Knostic documents an enforced allowlist, policy groups and audit logs of connections, blocks and drift events, and publishes prices (Kirin Team at $15 per user per month; AgentMesh tiers from free). Gaps on 16 Nov 2026: product documentation sits behind a sign-in, supported-client lists differ between pages, and named customers are none found. ## Strength and weakness for MCP Server Security **Strength:** Pairs pre-install scanning of MCP servers with a verdict (AgentMesh) and an endpoint control that validates and blocks unapproved MCP connections inline, with prices published (knostic-09, knostic-12, knostic-20). **Weakness:** Product documentation sits behind a sign-in and no named customers were found on 2026-11-16; supported-client lists differ between pages (knostic-22, knostic-29, knostic-06). ## Scores and evidence Knostic is 4th in the overall ranking at 61.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Knostic: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 6 of 10 | 25 | The vendor names Claude Code, Claude Desktop, Cursor, Windsurf and VS Code with Copilot and inventories MCP servers with publisher and transport; tools per server and remote-server discovery: none found. | [knostic-04](#fact-knostic-04), [knostic-06](#fact-knostic-06), [knostic-07](#fact-knostic-07), [knostic-03](#fact-knostic-03) | | Vetting before use | 7 of 10 | 20 | The vendor documents scanning servers for prompt injection and supply-chain threats, verdicts with findings, an enforced allowlist and published research, all from vendor sources. | [knostic-08](#fact-knostic-08), [knostic-09](#fact-knostic-09), [knostic-10](#fact-knostic-10), [knostic-11](#fact-knostic-11), [knostic-24](#fact-knostic-24) | | Runtime control of tool calls | 5 of 10 | 20 | The vendor claims inline enforcement and blocking of unapproved servers, and shows tool-call argument scoring; the example shows detection rather than a per-call block. | [knostic-12](#fact-knostic-12), [knostic-13](#fact-knostic-13), [knostic-14](#fact-knostic-14), [knostic-15](#fact-knostic-15) | | Deployment | 7 of 10 | 15 | Endpoint client by MDM or direct push in user space plus an IDE extension, with macOS, Windows and Linux beta and a minutes claim, all vendor-stated. | [knostic-16](#fact-knostic-16), [knostic-17](#fact-knostic-17), [knostic-06](#fact-knostic-06) | | Governance and audit | 6 of 10 | 5 | Audit logs, policy groups and SIEM streams are listed by plan, and MCP connections and blocks are logged; per-tool-call audit records: none found. | [knostic-18](#fact-knostic-18), [knostic-19](#fact-knostic-19), [knostic-20](#fact-knostic-20), [knostic-11](#fact-knostic-11) | | Transparency | 6 of 10 | 10 | Prices are published and open-source code and research are observed, but product documentation sits behind a sign-in. | [knostic-21](#fact-knostic-21), [knostic-22](#fact-knostic-22), [knostic-23](#fact-knostic-23), [knostic-24](#fact-knostic-24) | | Proof | 5 of 10 | 5 | Third-party reports confirm $14 million in funding and a directory listing; the analyst recognition is a vendor claim and named customers: none found. | [knostic-26](#fact-knostic-26), [knostic-28](#fact-knostic-28), [knostic-30](#fact-knostic-30), [knostic-29](#fact-knostic-29) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 4th | 61.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 4th | 61.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 5th | 63.0 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 4th | 63.0 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 6th | 58.0 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | joint 5th | 60.0 | ## Product, deployment, price and owner - Product assessed: Kirin (endpoint client and IDE extension for coding agents and MCP servers) and AgentMesh (scanning catalog for MCP servers, skills and extensions) - Deployment: Endpoint client (user space, MDM script or direct install) plus IDE extension for Cursor and VS Code - Time to deploy: Knostic says Kirin deploys in minutes - Price: Kirin Free (1 to 5 users), Team $15 per user per month, Enterprise by quote; AgentMesh Free, Premium $15, Pro $850 per user per month - Owner: Knostic (standalone company; no parent stated on its pages) - Website: [knostic.ai](https://www.knostic.ai) ## Facts and sources - **knostic-01** (vendor): Knostic's own About page lists Gadi Evron as co-founder and CEO and Sounil Yu as co-founder and CTO, and states no parent company or acquirer.Quote: "Sounil Yu Co-Founder and CTO"Source: [knostic.ai](https://www.knostic.ai/about), read 2026-11-16. - **knostic-02** (vendor): Knostic's pricing page frames two products for coding agents: Kirin for the agent on the endpoint and AgentMesh for the skills, MCP servers and extensions agents install.Quote: "Secure your AI coding agents with Kirin and the skills, MCP servers, and extensions they install with AgentMesh."Source: [knostic.ai](https://www.knostic.ai/pricing), read 2026-11-16. - **knostic-03** (vendor): Knostic lists cross-organization visibility into MCP servers, rules and hooks as a Team plan feature of Kirin.Quote: "Cross-org MCP, rules, hooks visibility"Source: [knostic.ai](https://www.knostic.ai/pricing), read 2026-11-16. - **knostic-04** (vendor): Knostic says Kirin keeps an inventory of every MCP server and validates each connection, flagging configuration drift.Quote: "Kirin keeps an inventory of every server, validates every connection, and flags configuration drift, while AgentMesh scans the servers themselves."Source: [knostic.ai](https://www.knostic.ai/solutions/mcp-security-posture-management), read 2026-11-16. - **knostic-05** (vendor): The Kirin product page lists the coding agents and IDEs it works with, including Codex, JetBrains and Gemini CLI.Quote: "Cursor, Claude Code, GitHub Copilot, Windsurf, Codex, JetBrains, Devin Desktop, and Gemini CLI."Source: [knostic.ai](https://www.knostic.ai/kirin), read 2026-11-16. - **knostic-06** (vendor): The pricing FAQ gives a narrower supported list (Claude Code CLI, Claude Desktop and Cowork, Cursor, Windsurf, VS Code with Copilot) on macOS and Windows with Linux in beta, and puts Codex and Gemini CLI on the roadmap.Quote: "Today, we support Claude Code CLI, Claude Desktop and Cowork, Cursor, Windsurf, and VS Code + GitHub Copilot, on macOS and Windows. Linux is in beta."Source: [knostic.ai](https://www.knostic.ai/pricing), read 2026-11-16. - **knostic-07** (vendor): A Kirin alert example on the product page records each MCP server's publisher and transport type (for example local stdio), and a dashboard example ranks MCP servers by usage; remote-server discovery is not described in words.Quote: "Publisher Unknown Type Local · stdio"Source: [knostic.ai](https://www.knostic.ai/kirin), read 2026-11-16. - **knostic-08** (vendor): Knostic says AgentMesh continuously discovers and scans agent skills, MCP servers and VS Code extensions for prompt injection and supply-chain threats.Quote: "AgentMesh continuously discovers, tracks, and scans AI agent skills, MCP servers, and VS Code extensions for prompt injection and supply-chain threats"Source: [knostic.ai](https://www.knostic.ai/agentmesh), read 2026-11-16. - **knostic-09** (vendor): Knostic says AgentMesh rates a server before use and Kirin blocks the connection to a scored-bad server before it is established.Quote: "AgentMesh scans the server and rates it. Kirin blocks the connection before it is established."Source: [knostic.ai](https://www.knostic.ai/solutions/mcp-security-posture-management), read 2026-11-16. - **knostic-10** (vendor): Knostic says every scanned item gets a dangerous, risky or clean verdict with the findings behind it, and items can be looked up by SHA-256.Quote: "Every item is marked dangerous, risky, or clean, with the scan findings behind the call."Source: [knostic.ai](https://www.knostic.ai/agentmesh), read 2026-11-16. - **knostic-11** (vendor): Knostic says Kirin maintains an up-to-date list of allowed MCP servers and validates every connection against it.Quote: "An up-to-date list of allowed MCP servers, with every connection validated in real time."Source: [knostic.ai](https://www.knostic.ai/solutions/mcp-security-posture-management), read 2026-11-16. - **knostic-12** (vendor): Knostic says Kirin inspects MCP connections and enforces policy inline before the agent acts.Quote: "It validates connections and blocks unsafe behavior in real time, inspecting dependencies, extensions, and MCP connections and enforcing policy inline, before the agent acts"Source: [knostic.ai](https://www.knostic.ai/kirin), read 2026-11-16. - **knostic-13** (vendor): Knostic says Kirin's IDE extension sits inside the coding agent and detects prompts, tool calls, file actions and MCP traffic.Quote: "Kirin's IDE extension sits inside the coding agent only (e.g., VS Code Copilot, Cursor, Windsurf), detecting prompts, tool calls, file actions, and MCP traffic."Source: [knostic.ai](https://www.knostic.ai/faq), read 2026-11-16. - **knostic-14** (vendor): A Kirin product-page example shows a Bash tool call scored against a threshold and flagged with outcome Detected, which shows argument inspection but in that example detection rather than a block.Quote: "Real-time tool-call analysis flagged this call"Source: [knostic.ai](https://www.knostic.ai/kirin), read 2026-11-16. - **knostic-15** (vendor): Knostic says Kirin blocks unapproved MCP servers in real time.Quote: "By enforcing secure configuration, maintaining the server inventory, validating connections, and blocking unapproved servers in real time."Source: [knostic.ai](https://www.knostic.ai/solutions/mcp-security-posture-management), read 2026-11-16. - **knostic-16** (vendor): Knostic says Kirin deploys in minutes by pushing an endpoint client directly or through MDM and starting in Monitor mode.Quote: "Knostic Kirin deploys in minutes. Managers provision the console, push the endpoint client to the endpoints (directly or via your existing MDM), and start in Monitor mode."Source: [knostic.ai](https://www.knostic.ai/pricing), read 2026-11-16. - **knostic-17** (vendor): Knostic says the Kirin client runs in user space without admin rights and is paired with an IDE extension for Cursor and VS Code.Quote: "The client runs in user-space (does not require admin)."Source: [knostic.ai](https://www.knostic.ai/pricing), read 2026-11-16. - **knostic-18** (vendor): Knostic says audit logs and per-endpoint activity logs are in-product on Team and above, with streams to a SIEM on Enterprise.Quote: "Audit logs and per-endpoint activity logs are available in-product on Team and above. On Enterprise, those streams pipe into your SIEM"Source: [knostic.ai](https://www.knostic.ai/pricing), read 2026-11-16. - **knostic-19** (vendor): Knostic lists policy groups, SIEM and Slack integrations and enterprise SSO in the Enterprise plan.Quote: "Policy groups Advanced contextual detections (add-on included) Enterprise SSO Integrations (SIEM, Slack)"Source: [knostic.ai](https://www.knostic.ai/pricing), read 2026-11-16. - **knostic-20** (vendor): Knostic says Kirin logs MCP connections, blocks and drift events.Quote: "Connections, blocks, and drift events logged for review."Source: [knostic.ai](https://www.knostic.ai/solutions/mcp-security-posture-management), read 2026-11-16. - **knostic-21** (vendor): Knostic publishes prices: Kirin Free for 1 to 5 users, Team at $15 per user per month, Enterprise by quote; AgentMesh Free, Premium $15 and Pro $850 per user per month.Quote: "Team $15 /mo, per user"Source: [knostic.ai](https://www.knostic.ai/pricing), read 2026-11-16. - **knostic-22** (none-found): The docs site docs.knostic.ai redirected to a GitBook sign-in page when read, so product documentation is not readable without sign-in.Source: [docs.knostic.ai](https://docs.knostic.ai), read 2026-11-16. - **knostic-23** (observed): Knostic publishes MCP-Scanner on GitHub under the MIT license, a research tool that finds publicly exposed MCP servers through Shodan (internet exposure, not servers configured on developer machines).Quote: "Advanced security research tool for discovering and analyzing Model Context Protocol (MCP) servers using the Shodan search engine."Source: [github.com](https://github.com/knostic/MCP-Scanner), read 2026-11-16. - **knostic-24** (vendor): On 2026-11-12, Knostic published research on an MCP package whose tool description tells the agent to obtain a GitHub token without asking the user; it was a static review and the package was not run.Quote: "In version 1.0.13, the description of its get_started tool contains a prompt injection: it tells the agent to obtain a GitHub token without asking the user."Source: [knostic.ai](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token), read 2026-11-16. - **knostic-25** (vendor): On 2026-10-26, Knostic published research on a repackaged Amazon SP-API MCP package that AgentMesh flagged for sending application credentials to a hardcoded IP.Quote: "AgentMesh flagged version 0.0.7"Source: [knostic.ai](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip), read 2026-11-16. - **knostic-26** (third-party): SecurityWeek reported that Knostic has raised a total of $14 million.Quote: "has raised a total of $14 million to build and market a"Source: [securityweek.com](https://www.securityweek.com/knostic-secures-11-million-to-rein-in-enterprise-ai-data-leakage-oversharing/), read 2026-11-16. - **knostic-27** (vendor): Knostic says it raised an additional $11 million, bringing total funding to $14.3 million.Quote: "we've raised an additional $11 million, bringing our total funding to $14.3 million"Source: [knostic.ai](https://www.knostic.ai/blog/ending-llm-oversharing-we-raised-11-million-to-secure-enterprise-ai), read 2026-11-16. - **knostic-28** (vendor): Knostic says Gartner named it a 2025 Cool Vendor in AI Cybersecurity Governance (the claim is on Knostic's own blog; the Gartner report itself was not read).Quote: "has been recognized as a 2025 Gartner® Cool Vendor in AI Cybersecurity Governance"Source: [knostic.ai](https://www.knostic.ai/blog/gartner-cool-vendor-recognition), read 2026-11-16. - **knostic-29** (none-found): None found on 2026-11-16: named customers on the Knostic home, about, pricing, Kirin and AgentMesh pages.Source: [knostic.ai](https://www.knostic.ai/), read 2026-11-16. - **knostic-30** (third-party): The CybersecTools directory lists Kirin AI Coding Safety as a Knostic product for AI coding assistants, agents and MCP servers (a standalone directory entry, not an analyst report).Quote: "Kirin AI Coding Safety is a security solution designed to protect AI coding assistants, AI agents, and Model Context Protocol servers."Source: [cybersectools.com](https://cybersectools.com/tools/kirin-ai-coding-safety), read 2026-11-16. ## News mentioning Knostic - [Knostic: MCP tool asks agents for a GitHub token](/news/knostic-mcp-tool-description-asks-agent-for-github-token/) - [Knostic flags a repackaged Amazon SP-API MCP server](/news/knostic-repackaged-amazon-sp-api-mcp-credential-redirect/) ## Related pages - [All vendor profiles](/vendors/) - [Harmonic Security profile](/vendors/harmonic-security/) - [Koi profile](/vendors/koi/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Koi MCP security: endpoint discovery through Cortex URL: https://mcpservercybersecurity.com/vendors/koi/ Updated: 2026-11-16 Koi's endpoint discovery now reaches customers through Palo Alto Networks, where third-party press reported a $400 million acquisition. Cortex documentation says the Cortex XDR agent discovers agentic software on the endpoint, including AI coding tools, MCP servers and IDE extensions, with remediation by policy, on Windows and macOS but not Linux, delivered by the XDR agent or a targeted sensor. Third-party reports confirmed Koi's scanning of extension stores and a malicious MCP server find. Gaps on 16 Nov 2026: named MCP clients, per-tool-call allow, block or approval with arguments inspected, a per-server MCP audit log, a time to deploy, a public price and readable research pages are none found. ## Strength and weakness for MCP Server Security **Strength:** Palo Alto Networks documents endpoint discovery of MCP servers, IDE extensions and packages with remediation by policy, on Windows and macOS (koi-13, koi-14, koi-19). **Weakness:** None found on 2026-11-16 for named MCP clients, per-tool-call argument inspection, a time to deploy, a public price or readable docs (koi-15, koi-18, koi-22, koi-26, koi-27). ## Scores and evidence Koi is 13th in the overall ranking at 39.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Koi: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 4 of 10 | 25 | Observed docs show discovery of MCP servers and IDE extensions on endpoints; named MCP clients read: none found. | [koi-13](#fact-koi-13), [koi-14](#fact-koi-14), [koi-15](#fact-koi-15) | | Vetting before use | 5 of 10 | 20 | Third-party reports confirm store scanning and a malicious MCP server find; MCP tool-description scanning or an enforced server registry: none found. | [koi-10](#fact-koi-10), [koi-11](#fact-koi-11), [koi-12](#fact-koi-12), [koi-09](#fact-koi-09) | | Runtime control of tool calls | 2 of 10 | 20 | Blocking is described in general terms; per-tool-call allow, block or approval with arguments inspected: none found. | [koi-16](#fact-koi-16), [koi-17](#fact-koi-17), [koi-18](#fact-koi-18) | | Deployment | 6 of 10 | 15 | Observed docs show delivery by the XDR agent or a targeted sensor on Windows and macOS; Linux and a time to deploy: none found. | [koi-19](#fact-koi-19), [koi-20](#fact-koi-20), [koi-22](#fact-koi-22) | | Governance and audit | 2 of 10 | 5 | Governance is stated generally; a per-server MCP audit log or SIEM export: none found. | [koi-23](#fact-koi-23), [koi-24](#fact-koi-24), [koi-25](#fact-koi-25) | | Transparency | 1 of 10 | 10 | Research was reported by a third party but the posts no longer load; public docs, price and open-source code: none found. | [koi-26](#fact-koi-26), [koi-27](#fact-koi-27), [koi-28](#fact-koi-28), [koi-09](#fact-koi-09) | | Proof | 8 of 10 | 5 | Third-party press confirms a $400 million acquisition and prior funding, with named customer logos; the analyst report is listed by the vendor only. | [koi-05](#fact-koi-05), [koi-06](#fact-koi-06), [koi-07](#fact-koi-07), [koi-08](#fact-koi-08) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 13th | 39.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 13th | 41.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 12th | 45.5 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | joint 11th | 41.5 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 14th | 34.0 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 14th | 34.0 | ## Product, deployment, price and owner - Product assessed: Cortex Agentic Endpoint Security (Koi), standalone or as a Cortex XDR module - Deployment: Endpoint sensor: Cortex XDR agent, or a Cortex targeted sensor alongside third-party EDR - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 - Owner: Palo Alto Networks - Website: koi.ai ## Facts and sources - **koi-01** (vendor): Palo Alto Networks announced it completed the acquisition of Koi; the press release is dated April 14, 2026.Quote: "today announced the completion of its acquisition of Koi. With this acquisition, Palo Alto Networks defines a new category of protection: Agentic Endpoint Security (AES)."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/company/press/2026/palo-alto-networks-completes-acquisition-of-koi-to-secure-the-agentic-endpoint), read 2026-11-16. - **koi-02** (vendor): Palo Alto Networks blog dated Feb 17, 2026 also announces the completed acquisition (the date differs from the April 14 press release; both are recorded as published).Quote: "Palo Alto Networks has officially completed the acquisition of"Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/blog/2026/02/securing-the-agentic-endpoint/), read 2026-11-16. - **koi-03** (observed): The koi.ai domain now redirects (HTTP 301) to the Palo Alto Networks Cortex Agentic Endpoint Security page, which carries the Koi product.Quote: "Secure the modern software stack with Koi. Gain visibility, context, and real-time protection over extensions, MCPs, and autonomous AI agent endpoints."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/cortex/agentic-endpoint-security), read 2026-11-16. - **koi-04** (vendor): Palo Alto Networks says Koi stays available as a standalone offering and as a Cortex XDR module.Quote: "Koi's capabilities will also remain available as a standalone offering, allowing every customer to benefit from industry leading agentic endpoint security capabilities alongside their existing EDR solutions."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/company/press/2026/palo-alto-networks-completes-acquisition-of-koi-to-secure-the-agentic-endpoint), read 2026-11-16. - **koi-05** (third-party): Globes reported on 2026-02-17 that Palo Alto Networks would buy Koi Security, with the price estimated at $400 million, and named the three founders.Quote: "No financial details were disclosed but estimates are that Palo Alto will pay $400 million for the company."Source: [en.globes.co.il](https://en.globes.co.il/en/article-palo-alto-networks-to-buy-israeli-co-koi-security-1001535257), read 2026-11-16. - **koi-06** (third-party): Globes, citing PitchBook, reported Koi had raised $48 million over two rounds and named its investors.Quote: "Koi has raised only $48 million over two rounds, according to PitchBook"Source: [en.globes.co.il](https://en.globes.co.il/en/article-palo-alto-networks-to-buy-israeli-co-koi-security-1001535257), read 2026-11-16. - **koi-07** (vendor): The Palo Alto Networks AES page shows named customers FHLBank, Jump Trading and Fireblocks (logo list under Customers).Quote: "FHLBank Jump Trading Fireblocks"Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/cortex/agentic-endpoint-security), read 2026-11-16. - **koi-08** (vendor): Palo Alto Networks lists a Gartner report titled "Confronting the Risks of Shadow AI With Endpoint Application Control" among its AES resources; the report itself needs sign-in.Quote: "Confronting the Risks of Shadow AI With Endpoint Application Control"Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/cortex/agentic-endpoint-security), read 2026-11-16. - **koi-09** (third-party): On 2025-09-29 The Hacker News reported that Koi Security found the first malicious MCP server in the wild, a fake postmark-mcp npm package.Quote: ""This is the world's first sighting of a real-world malicious MCP server."Source: [thehackernews.com](https://thehackernews.com/2025/09/first-malicious-mcp-server-found.html), read 2026-11-16. - **koi-10** (vendor): Palo Alto Networks says Koi continuously analyzes the risk of AI software and agents (three pillars: see, understand risk, control).Quote: "Continuously analyze and understand the intent and risk level of all software and AI agents."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/blog/2026/02/securing-the-agentic-endpoint/), read 2026-11-16. - **koi-11** (third-party): Globes reports Koi scans extension and package stores (VS Code Marketplace, Edge, Chrome, Firefox, NPM, Homebrew) with an LLM and agent based engine.Quote: "Koi's engine scans application and extension stores such as Microsoft's Visual Studio (VSCode Marketplace), Microsoft's Edge, the Chrome store (Google), the Firefox store, NPM and Homebrew."Source: [en.globes.co.il](https://en.globes.co.il/en/article-palo-alto-networks-to-buy-israeli-co-koi-security-1001535257), read 2026-11-16. - **koi-12** (vendor): Palo Alto Networks says AES offers out-of-the-box policies and one-click removal of insecure packages, extensions and AI tools.Quote: "Deploy out-of-the-box policies that govern agent interactions and allow one-click removal of insecure packages, extensions, and AI tools."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/cortex/agentic-endpoint-security), read 2026-11-16. - **koi-13** (vendor): The Cortex AES datasheet says it discovers shadow AI, browser extensions, IDE plugins, MCP servers and other unmanaged software on endpoints.Quote: "Discover shadow AI, browser extensions, IDE plugins, MCP servers, and other unmanaged software across endpoints"Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/resources/datasheets/cortex-agentic-endpoint-security-at-a-glance), read 2026-11-16. - **koi-14** (observed): Cortex documentation says the Cortex XDR agent discovers agentic software on the endpoint, including AI coding tools, MCP servers and IDE extensions.Quote: "the Cortex XDR agent discovers agentic software running on the endpoint and remediates risks based on your AES policy."Source: [cortex-docs.paloaltonetworks.com](https://cortex-docs.paloaltonetworks.com/cortex-xsiam/protect-your-endpoints/endpoint-security/endpoint-protection/endpoint-protection-capabilities), read 2026-11-16. - **koi-15** (none-found): None found on 2026-11-16: a list of named MCP clients (Claude Desktop, Cursor, VS Code, Claude Code, others) that Koi reads MCP configuration from, on the AES page, datasheet, blog posts and Cortex docs; the Koi docs at docs.koi.ai redirect to a sign-in page.Source: [docs.koi.ai](https://docs.koi.ai/), read 2026-11-16. - **koi-16** (vendor): Palo Alto Networks says AES stops risky downloads and rogue AI agent actions before they cause damage.Quote: "stop risky downloads and rogue AI agent actions before they cause damage."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/resources/datasheets/cortex-agentic-endpoint-security-at-a-glance), read 2026-11-16. - **koi-17** (vendor): Palo Alto Networks says Cortex XDR One controls an agent's actions at the source and enforces OS-level hard boundaries (blog dated Oct 19, 2026).Quote: "XDR One is designed to control an agent's actions at the source, enforce hard boundaries at the operating system level"Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/blog/2026/10/reinventing-endpoint-security-with-cortex-xdr-one/), read 2026-11-16. - **koi-18** (none-found): None found on 2026-11-16: whether AES inspects MCP tool-call arguments and lets an admin allow, block or require approval per tool call; pages describe policy and blocking in general terms only.Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/cortex/agentic-endpoint-security), read 2026-11-16. - **koi-19** (observed): Cortex documentation lists AES as supported on Windows and macOS (not Linux) and delivered by the Cortex XDR agent.Quote: "Windows macOS"Source: [cortex-docs.paloaltonetworks.com](https://cortex-docs.paloaltonetworks.com/cortex-xsiam/protect-your-endpoints/endpoint-security/endpoint-protection/endpoint-protection-capabilities), read 2026-11-16. - **koi-20** (vendor): Palo Alto Networks says AES can run alongside third-party EDR using the Cortex targeted sensor, or natively with XDR in a single sensor.Quote: "Organizations have the flexibility to deploy AES alongside existing third-party EDR environments using the Cortex targeted sensor."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/blog/2026/10/reinventing-endpoint-security-with-cortex-xdr-one/), read 2026-11-16. - **koi-21** (vendor): Palo Alto Networks release notes say Cortex XDR 6.0 (November 2026) adds Agentic Endpoint Security and targeted sensors.Quote: "Cortex XDR 6.0 adds Agentic Endpoint Security, flexible Targeted Sensors, and on-endpoint SLMs."Source: [cortex-docs.paloaltonetworks.com](https://cortex-docs.paloaltonetworks.com/), read 2026-11-16. - **koi-22** (none-found): None found on 2026-11-16: a published time to deploy for AES or Koi on developer machines, on the AES page, datasheet, blog and Cortex docs.Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/cortex/agentic-endpoint-security), read 2026-11-16. - **koi-23** (vendor): Palo Alto Networks says AES policies are enforced with least-privilege guardrails; the guide names proactive governance and control as a pillar.Quote: "Enforce least-privilege guardrails to maximize security and innovation"Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/resources/guides/definitive-guide-to-aes), read 2026-11-16. - **koi-24** (vendor): Palo Alto Networks says XDR One gives centralized visibility, management and investigation in one interface.Quote: "XDR One integrates Cortex XDR and Cortex Agentic Endpoint Security with a single user interface and sensor, providing centralized visibility, management and investigation."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/blog/2026/10/reinventing-endpoint-security-with-cortex-xdr-one/), read 2026-11-16. - **koi-25** (none-found): None found on 2026-11-16: a documented per-server audit log of MCP tool calls or SIEM export for AES; the Koi docs at docs.koi.ai require sign-in.Source: [docs.koi.ai](https://docs.koi.ai/), read 2026-11-16. - **koi-26** (observed): The Koi documentation site (docs.koi.ai), linked from Cortex docs, redirects to a sign-in page, so it is not readable without an account.Quote: "Learn more about Agentic Endpoint Security with Koi (https://docs.koi.ai/)."Source: [cortex-docs.paloaltonetworks.com](https://cortex-docs.paloaltonetworks.com/cortex-xsiam/protect-your-endpoints/endpoint-security/endpoint-protection/endpoint-protection-capabilities), read 2026-11-16. - **koi-27** (none-found): None found on 2026-11-16: a published price for Koi or Cortex AES; the AES page offers a demo request form.Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/cortex/agentic-endpoint-security), read 2026-11-16. - **koi-28** (none-found): None found on 2026-11-16: Koi's own research posts (for example the postmark-mcp post at koi.ai/blog) no longer load; koi.ai/blog redirects to the Palo Alto Networks AES page. No open-source repository for the product was found (github.com/koidex is an unrelated personal account).Source: [koi.ai](https://www.koi.ai/blog), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Knostic profile](/vendors/knostic/) - [Lasso Security profile](/vendors/lasso-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Lasso Security MCP security: discovery, scanning and call monitoring URL: https://mcpservercybersecurity.com/vendors/lasso-security/ Updated: 2026-11-16 Lasso Security says it discovers MCP servers across the organization and inventories Claude Code, Claude Desktop, Cursor, Windsurf, Codex and OpenCode. It scores servers on permissions and descriptions, scans tool descriptions for hidden instructions, and inspects every tool call before execution with flag, block or masking options. Lasso says it keeps audit trails of every MCP connection and tool call, exports logs to a SIEM and applies role-based control per team and server. It deploys through hooks pushed by an enterprise management platform, with no agent, and an open-source gateway is published. Gaps on 16 Nov 2026: supported operating systems, deploy time, VS Code coverage, public documentation and a price are none found; capability statements are the vendor's. ## Strength and weakness for MCP Server Security **Strength:** Documents MCP server discovery with risk scores, tool-description scanning and tool-call monitoring with SIEM export (lasso-security-19, lasso-security-21, lasso-security-25). **Weakness:** No operating systems, deploy time, VS Code coverage or public docs found (lasso-security-30, lasso-security-52; none found on 2026-11-16). ## Scores and evidence Lasso Security is 7th in the overall ranking at 56.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Lasso Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 6 of 10 | 25 | The vendor names Claude Code, Claude Desktop, Cursor, Windsurf, Codex and OpenCode and inventories MCP servers; vendor claims only. | [lasso-security-19](#fact-lasso-security-19), [lasso-security-13](#fact-lasso-security-13), [lasso-security-45](#fact-lasso-security-45) | | Vetting before use | 6 of 10 | 20 | The vendor scores servers on permissions and descriptions and scans descriptions for hidden instructions; an enforced approved list: none found. | [lasso-security-20](#fact-lasso-security-20), [lasso-security-21](#fact-lasso-security-21) | | Runtime control of tool calls | 6 of 10 | 20 | The vendor claims inspecting every tool call before execution on the device with flag, block or masking; vendor claims only. | [lasso-security-12](#fact-lasso-security-12), [lasso-security-22](#fact-lasso-security-22), [lasso-security-28](#fact-lasso-security-28), [lasso-security-29](#fact-lasso-security-29) | | Deployment | 4 of 10 | 15 | Hooks deployed by the enterprise management platform with no agent; supported OS and a measured time: none found. | [lasso-security-26](#fact-lasso-security-26), [lasso-security-29](#fact-lasso-security-29), [lasso-security-30](#fact-lasso-security-30) | | Governance and audit | 7 of 10 | 5 | The vendor claims role-based control per team and server, audit trails of every tool call and SIEM export, all vendor-stated. | [lasso-security-24](#fact-lasso-security-24), [lasso-security-25](#fact-lasso-security-25) | | Transparency | 4 of 10 | 10 | Research and an observed open-source gateway are published; a price and public documentation: none found. | [lasso-security-31](#fact-lasso-security-31), [lasso-security-49](#fact-lasso-security-49), [lasso-security-51](#fact-lasso-security-51), [lasso-security-52](#fact-lasso-security-52) | | Proof | 7 of 10 | 5 | Third-party press confirms funding and names customers; analyst placement is a vendor claim. | [lasso-security-02](#fact-lasso-security-02), [lasso-security-03](#fact-lasso-security-03), [lasso-security-04](#fact-lasso-security-04) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 7th | 56.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 8th | 56.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 8th | 54.0 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 6th | 58.5 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 5th | 58.5 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 4th | 61.5 | ## Product, deployment, price and owner - Product assessed: Lasso MCP Security (MCP server discovery, risk scoring, tool-call monitoring) and open-source MCP Gateway - Deployment: Agentless (read-only CI and cloud integrations), API, SDK, gateway or firewall integration, browser extension, and Claude Code lifecycle hooks via enterprise management - Time to deploy: Lasso says start in hours, not weeks, and usage control in minutes - Price: none found on 2026-11-16 - Owner: Lasso Security Ltd. - Website: [lasso.security](https://www.lasso.security) ## Facts and sources - **lasso-security-01** (vendor): Lasso says its AI Security Platform gives enterprises visibility, control and protection across AI models, agents and apps.Quote: "Lasso's AI Security Platform gives enterprises visibility, control, and protection across AI models, agents, and apps."Source: [lasso.security](https://www.lasso.security/), read 2026-11-16. - **lasso-security-02** (third-party): SiliconANGLE reports that Lasso Security Ltd. was founded in 2023 by Schulman and Dror and has raised more than $37 million in total, including a $30 million round announced on 2026-09-02 led by ClearSky Advisors.Quote: "Investors have now put more than $37 million into the company."Source: [siliconangle.com](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/), read 2026-11-16. - **lasso-security-03** (third-party): SiliconANGLE names BMW, eToro, Fiverr and the U.S. Department of Homeland Security as Lasso customers (reported 2026-09-02).Quote: "BMW AG is a customer, and so are eToro Group Ltd., Fiverr International Ltd. and the U.S. Department of Homeland Security."Source: [siliconangle.com](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/), read 2026-11-16. - **lasso-security-04** (vendor): On 2026-09-15 Lasso said Latio Tech's 2026 AI Security Market Report named it a platform leader (vendor's own account of a third-party report).Quote: "In the report, Lasso is named as one of the major Platform Leaders."Source: [lasso.security](https://www.lasso.security/blog/lasso-named-a-leader-in-latios-2026-ai-security-market-report), read 2026-11-16. - **lasso-security-05** (third-party): On 2026-09-01 TechCentral (a promoted article by NEWORDER) said NEWORDER is Lasso's official partner in South Africa and that Lasso is certified to ISO 27001 and SOC 2.Quote: "Lasso Security was founded in Tel Aviv in 2023, is certified to ISO 27001 and SOC 2, and has customers across Europe, the US and Israel."Source: [techcentral.co.za](https://techcentral.co.za/neworder-lasso-security-ai-agent-controls/285558/), read 2026-11-16. - **lasso-security-12** (vendor): Lasso says it inspects every coding-agent tool call before execution and flags or blocks anything outside the permitted scope.Quote: "Lasso inspects every tool call before it is executed."Source: [lasso.security](https://www.lasso.security/use-cases/ai-coding-assistants), read 2026-11-16. - **lasso-security-13** (vendor): Lasso says its coding-assistant coverage addresses agents that read codebases, execute shell commands and call MCP servers (Claude Code, Cursor, Codex, OpenCode).Quote: "Claude Code, Cursor, and Codex read your codebase, executes shell commands, calls MCPs, and takes actions on your behalf."Source: [lasso.security](https://www.lasso.security/use-cases/ai-coding-assistants), read 2026-11-16. - **lasso-security-19** (vendor): Lasso says it discovers MCP servers across the organization and inventories Claude Code, Claude Desktop, Cursor, Windsurf and custom agent connections.Quote: "Automatically discover all MCP servers connected across your organization. Inventory Claude Code and Desktop, Cursor, Windsurf, and custom agent connections."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-20** (vendor): Lasso says it gives each MCP server a risk score from permissions, actions and description, and lets teams manage or block high-risk servers.Quote: "Get a risk score for each MCP according to its permissions, actions, description, and more. If an MCP server triggers a high risk score, manage or block it instantly."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-21** (vendor): Lasso says it scans MCP tool descriptions for hidden instructions and blocks prompt injection at the connection layer.Quote: "Scan MCP server tool descriptions for hidden instructions. Detect and block prompt injection attempts at the connection layer."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-22** (vendor): Lasso says it monitors every MCP tool call in real time and can mask sensitive data before it reaches external servers.Quote: "Monitor every MCP tool call in real-time to identify indirect prompt injection, memory poisoning, and other AI threats or attack techniques."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-23** (vendor): Lasso says it detects PII, API keys and credentials shared through MCP tool calls and masks them before they reach external servers.Quote: "Detect and monitor in real-time if PII, API keys, credentials, or any other sensitive data is shared through MCP tool calls. Mask sensitive contents before they reach external servers."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-24** (vendor): Lasso says it enforces role-based control over which users and teams can connect to which MCP servers.Quote: "Control which users and teams can connect to which MCP servers. Enforce role-based permissions that allow approved use cases while blocking unauthorized connections."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-25** (vendor): Lasso says it keeps audit trails of every MCP connection and tool call and exports logs to a SIEM.Quote: "Maintain complete audit trails of every MCP connection and tool call. Export logs to your SIEM."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-26** (vendor): Lasso says it governs AI activity on employee endpoints agentlessly, so visibility into AI tools, coding agents and MCP connections starts on day one.Quote: "Lasso governs AI activity on employee endpoints through an agentless approach, so security teams get visibility into AI tools, coding agents, and MCP connections from day one"Source: [lasso.security](https://www.lasso.security/blog/securing-ai-on-endpoints-how-to-govern-ai-tools-coding-agents-and-mcp-connections), read 2026-11-16. - **lasso-security-27** (vendor): Lasso says teams with EDR or MDM tooling can connect Lasso to that stack to extend endpoint coverage.Quote: "Teams that already run EDR or MDM tooling can connect Lasso to that existing stack, extending endpoint coverage through infrastructure they already maintain"Source: [lasso.security](https://www.lasso.security/blog/securing-ai-on-endpoints-how-to-govern-ai-tools-coding-agents-and-mcp-connections), read 2026-11-16. - **lasso-security-28** (vendor): Lasso says policy enforcement happens at the execution layer on the device itself.Quote: "Policy enforcement happens at the execution layer on the device itself, so protection holds even when an agent's activity would not otherwise pass through a monitored network path."Source: [lasso.security](https://www.lasso.security/blog/securing-ai-on-endpoints-how-to-govern-ai-tools-coding-agents-and-mcp-connections), read 2026-11-16. - **lasso-security-29** (vendor): Lasso says it attaches to Claude Code, Cursor and Codex through lifecycle hooks deployed by the enterprise management platform, with no agent to install.Quote: "Lasso deploys through the enterprise management platform, so there is no agent to install and no code to change."Source: [lasso.security](https://www.lasso.security/use-cases/ai-coding-assistants), read 2026-11-16. - **lasso-security-30** (none-found): None found on 2026-11-16: supported operating systems and a time to deploy for MCP discovery on developer machines, searched on the MCP security, coding assistants and endpoint pages.Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-31** (observed): Lasso publishes an open-source MCP gateway (MIT license) that intercepts requests and responses to sanitize sensitive information via plugins.Quote: "A plugin-based gateway that orchestrates other MCPs and allows developers to build upon it enterprise-grade agents."Source: [github.com](https://github.com/lasso-security/mcp-gateway), read 2026-11-16. - **lasso-security-36** (vendor): Lasso says AI usage control deploys through a browser extension, proxy integrations and data integrations, and can be running in minutes.Quote: "Enable full AI usage control in minutes through Lasso’s lightweight deployment paths like our browser extension, proxy integrations, and data integrations."Source: [lasso.security](https://www.lasso.security/platform/ai-usage-control), read 2026-11-16. - **lasso-security-45** (vendor): On 2026-09-29 Lasso said its Claude Compliance API integration is generally available, bringing Claude Enterprise users, MCP servers and skills into its inventory.Quote: "Today we're launching Lasso's integration with Claude's Compliance API, and it is generally available starting now."Source: [lasso.security](https://www.lasso.security/blog/bringing-claude-enterprise-activity-into-lasso-with-claudes-compliance-api), read 2026-11-16. - **lasso-security-47** (vendor): Lasso says it runs from the Lasso tenant and starts in hours, not weeks, with no agents or code changes.Quote: "No agents. No code changes. No source code access. Run from the Lasso tenant and start in hours, not weeks."Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-49** (vendor): On 2026-09-17 Lasso published research on how LLM text watermarking changes refusals and agent tool calling.Quote: "We find that it does, in both model refusal behavior and agent tool calling."Source: [lasso.security](https://www.lasso.security/blog/the-provenance-tax-understanding-the-impact-of-llm-watermarking-on-ai-agent-behavior), read 2026-11-16. - **lasso-security-50** (vendor): On 2026-09-09 Lasso published a technical post on LEAP, its CPU-based guardrail, with benchmark claims.Quote: "Introducing LEAP: CPU-based AI Security Guardrails with GPU-Class Accuracy"Source: [lasso.security](https://www.lasso.security/blog/introducing-leap-cpu-based-ai-security-guardrails-with-gpu-class-accuracy), read 2026-11-16. - **lasso-security-51** (none-found): None found on 2026-11-16: a published price; the pricing URL returns a 404 page and the site routes to Book a Demo.Source: [lasso.security](https://www.lasso.security/pricing), read 2026-11-16. - **lasso-security-52** (none-found): None found on 2026-11-16: public product documentation; docs.lasso.security loads only a sign-in dashboard titled Lasso Security - Dashboard.Source: [docs.lasso.security](https://docs.lasso.security), read 2026-11-16. - **lasso-security-53** (vendor): Lasso says its detection covers indirect prompt injection in files, web responses and MCP outputs before they reach the coding agent.Quote: "detect injected payloads or hidden instructions in files, web responses, and MCP outputs before they reach the coding agent’s decision layer"Source: [lasso.security](https://www.lasso.security/use-cases/ai-coding-assistants), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Koi profile](/vendors/koi/) - [Microsoft profile](/vendors/microsoft/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Microsoft MCP security: Defender for Endpoint URL: https://mcpservercybersecurity.com/vendors/microsoft/ Updated: 2026-11-16 Microsoft's Defender for Endpoint inventories local AI agents and the MCP servers each has configured on Windows and macOS, using existing device onboarding with Intune or PowerShell. Its support matrix names Claude Code, Claude Desktop, Cursor, Codex CLI, Gemini CLI, GitHub Copilot, Windsurf and Kiro, and a third-party report confirmed the breadth of local agent discovery. Runtime protection that inspects the agent loop and blocks risky activity is a public preview, and rules for MCP tools onboarded to Agent 365 can audit or block calls. Agent 365 is listed at $15.00 per user per month. Gaps on 16 Nov 2026: scanning of server code or tool descriptions, an enforced approved-server registry and SIEM export outside Microsoft are none found. ## Strength and weakness for MCP Server Security **Strength:** Defender for Endpoint inventories many named local agents and their local and remote MCP servers on Windows and macOS with no extra deployment (microsoft-14, microsoft-15, microsoft-12). **Weakness:** None found on 2026-11-16 for scanning MCP servers before use or an enforced approved list; local runtime protection is preview and covers few agents (microsoft-27, microsoft-26). ## Scores and evidence Microsoft is 8th in the overall ranking at 54.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Microsoft: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 8 of 10 | 25 | The vendor names many local clients (Claude Code, Claude Desktop, Cursor, Copilot, Windsurf and others) and lists MCP servers per agent with declared tools; a third-party report confirms broad local agent discovery. Discovery is documented for Windows and macOS only. | [microsoft-14](#fact-microsoft-14), [microsoft-15](#fact-microsoft-15), [microsoft-16](#fact-microsoft-16), [microsoft-17](#fact-microsoft-17), [microsoft-65](#fact-microsoft-65) | | Vetting before use | 2 of 10 | 20 | The vendor mentions controlling which tools agents connect to and approving agents before publishing; scanning of server code or tool descriptions and an enforced approved-server registry: none found. | [microsoft-68](#fact-microsoft-68), [microsoft-69](#fact-microsoft-69), [microsoft-27](#fact-microsoft-27) | | Runtime control of tool calls | 4 of 10 | 20 | On-endpoint blocking of tool calls is in public preview on Windows with documented limits, so it counts only as a mention; the non-preview control is rules that can block tool calls for MCP tools onboarded to Agent 365, not shown on the machine. Generally available on-machine enforcement with arguments inspected: none found. | [microsoft-28](#fact-microsoft-28), [microsoft-19](#fact-microsoft-19), [microsoft-23](#fact-microsoft-23), [microsoft-26](#fact-microsoft-26) | | Deployment | 6 of 10 | 15 | Deployment rides on existing device onboarding with Intune or PowerShell and an OS list is published; a time to deploy: none found. | [microsoft-12](#fact-microsoft-12), [microsoft-13](#fact-microsoft-13), [microsoft-25](#fact-microsoft-25) | | Governance and audit | 5 of 10 | 5 | Inventory and advanced hunting cover MCP servers and tools, and rules can audit or block tool calls; per-team policy and export to an outside SIEM: none found. | [microsoft-16](#fact-microsoft-16), [microsoft-17](#fact-microsoft-17), [microsoft-28](#fact-microsoft-28), [microsoft-68](#fact-microsoft-68), [microsoft-39](#fact-microsoft-39) | | Transparency | 8 of 10 | 10 | Documentation is public and an open-source MCP gateway is observed on GitHub; a price is published for one component but none found for the main product. | [microsoft-52](#fact-microsoft-52), [microsoft-49](#fact-microsoft-49), [microsoft-51](#fact-microsoft-51), [microsoft-53](#fact-microsoft-53) | | Proof | 6 of 10 | 5 | Third-party press covered the launch of a large established vendor; named customers and analyst coverage for this product: none found. | [microsoft-64](#fact-microsoft-64), [microsoft-58](#fact-microsoft-58) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 8th | 54.5 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 7th | 56.5 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 4th | 63.5 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 10th | 44.0 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | joint 10th | 49.0 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 7th | 54.0 | ## Product, deployment, price and owner - Product assessed: Defender for Endpoint local AI agent discovery (MCP server configuration) and AI agent runtime protection, Agent 365 real-time protection for MCP tools - Deployment: Cloud-managed Defender for Endpoint client onboarding (Intune or PowerShell for runtime protection); API-based Content Safety; Purview browser extension; Entra and Defender portal controls - Time to deploy: none found on 2026-11-16 (discovery starts automatically after device onboarding; Purview policies need at least 24 hours) - Price: Agent 365 is listed at $15.00 per user per month, paid yearly; Content Safety Standard shows a dash; Defender for Endpoint and Cloud Apps prices none found on 2026-11-16 on readable pages. - Owner: Microsoft Corporation - Website: learn.microsoft.com ## Facts and sources - **microsoft-11** (vendor): Microsoft says Defender for Endpoint discovers supported local AI agents on Windows devices, including coding agents, IDE extensions, desktop assistants and agent platforms, and lists them in an AI agent inventory.The exact quote is on the source page.Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint), read 2026-11-16. - **microsoft-12** (vendor): Microsoft says local agent discovery needs no extra deployment beyond onboarding to Defender for Endpoint and starts automatically; Plan 2 is the minimum license.Quote: "You don't need additional deployment, configuration, or scripts beyond the device onboarding requirements. If the device meets all of the prerequisites, agent discovery begins automatically."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents), read 2026-11-16. - **microsoft-13** (vendor): Microsoft lists Windows, macOS, Linux, Android and iOS as Defender for Endpoint operating systems; local agent discovery is documented for Windows and macOS.Quote: "Microsoft Defender for Endpoint supports the following operating systems: Windows, macOS, Linux, Android, and iOS."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/microsoft-defender-endpoint), read 2026-11-16. - **microsoft-14** (vendor): Microsoft's support matrix names the local agents discovered, such as Claude Code, Claude Desktop, Cursor, Codex CLI, Gemini CLI, GitHub Copilot, Windsurf and Kiro, and marks which also report MCP server configuration.Quote: "Discovers MCP column indicates that Defender reports remote MCP server data in McpServers and local MCP server data in RawAgentInfo.localAgentMetadata.localMcps"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-support-matrix), read 2026-11-16. - **microsoft-15** (vendor): On 2026-10 release notes, Microsoft says local AI agent discovery is generally available on Windows and macOS and reports each agent's configured MCP servers.Quote: "Discover supported local AI agents and their configured MCP servers on onboarded Windows and macOS devices"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint), read 2026-11-16. - **microsoft-16** (vendor): Microsoft says the AI agent inventory lists, per local agent, the remote and local MCP servers with name, type and endpoint.Quote: "MCP servers lists the MCP servers configured for the agent, with the name, type, and endpoint of each server."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents), read 2026-11-16. - **microsoft-17** (vendor): Microsoft says local agent data can be queried in advanced hunting, including an AgentsInfo table that carries MCP servers and declared tools.Quote: "Use advanced hunting to proactively investigate local AI agent presence, understand how agents are configured, and identify the agents and users that carry the most risk."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents), read 2026-11-16. - **microsoft-18** (vendor): Microsoft says the agent risk level, indicators and recommendations for discovered local agents need Microsoft 365 E7 or Agent 365 together with Defender for Endpoint Plan 2.Quote: "View the risk level, risk indicators, and security recommendations for discovered local AI agents"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents), read 2026-11-16. - **microsoft-19** (vendor): Microsoft says Defender for Endpoint runtime protection inspects the agent loop on endpoints and can block risky activity before it executes; this is a public preview feature.Quote: "Microsoft Defender inspects the agent loop (user prompts, tool calls, and tool responses) and can block risky activity before it executes"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint), read 2026-11-16. - **microsoft-20** (vendor): Microsoft documents agent-native event inspection for agents that expose vendor event interfaces (Claude Code, Codex CLI, GitHub Copilot CLI) and network inspection for others.Quote: "Agents such as Claude Code, Codex CLI, and GitHub Copilot CLI expose these interfaces, and Defender uses them to inspect agent activity"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-23** (vendor): Microsoft lists preview limits: network inspection skips agents using certificate pinning or HTTP/3, and Claude Code may keep processing a tool response in Block mode after a post-tool-response block.Quote: "Network inspection doesn't support agents that use certificate pinning or HTTP/3."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-25** (vendor): Microsoft says runtime protection is configured in audit or block mode through local PowerShell or organization-wide Intune.Quote: "For local PowerShell and organization-wide Microsoft Intune configuration options"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-26** (vendor): The support matrix lists runtime protection per agent on Windows (preview), including Claude Code, GitHub Copilot CLI and app, Claude Desktop Code, VS Code Copilot and Claude Code extensions, plus OpenClaw, Ollama Desktop and ChatGPT Desktop via network inspection.Quote: "Network inspection covers agents that communicate with LLMs through well-known inference services. It doesn't cover agents that use certificate pinning or HTTP/3."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-support-matrix), read 2026-11-16. - **microsoft-27** (none-found): None found on 2026-11-16: no Microsoft page for Defender for Endpoint describes scanning of MCP server code or tool descriptions or an enforced approved-server registry before use; checked the local agent discovery, support matrix and runtime protection pages.Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents), read 2026-11-16. - **microsoft-28** (vendor): Microsoft Defender integrates with Agent 365 so that real-time protection evaluates tool calls and responses for Work IQ MCP and customer MCP tools onboarded to Agent 365; the default rule audits and custom rules can block before execution.Quote: "custom rules can block matching actions before they execute"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-cloud-apps/ai-agent-protection), read 2026-11-16. - **microsoft-37** (vendor): Microsoft says Purview keeps a record of agent activity for audits and admins control which tools agents can connect to.Quote: "Microsoft Purview keeps a record of agent activity, so you're ready for audits."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/microsoft-agent-365/overview), read 2026-11-16. - **microsoft-39** (observed): Microsoft publishes an open-source MCP Gateway, a reverse proxy and management layer for MCP servers on Kubernetes with authorization and lifecycle management (an MIT-licensed GitHub repo, last pushed 2026-11-13).Quote: "reverse proxy and management layer for [Model Context Protocol (MCP)](https://modelcontextprotocol.io/introduction) servers"Source: [github.com](https://github.com/microsoft/mcp-gateway), read 2026-11-16. - **microsoft-49** (vendor): Microsoft lists Agent 365 at $15.00 per user per month, paid yearly, on its product page.Quote: "$15.00 user/month, paid yearly (Annual commitment)"Source: [microsoft.com](https://www.microsoft.com/en-us/microsoft-agent-365), read 2026-11-16. - **microsoft-50** (vendor): Microsoft's Azure AI Content Safety pricing page shows a free tier of 5,000 text records per month and a Standard tier whose price displays as a dash.The exact quote is on the source page.Source: [azure.microsoft.com](https://azure.microsoft.com/en-us/pricing/details/content-safety/), read 2026-11-16. - **microsoft-51** (none-found): None found on 2026-11-16: no price for Defender for Cloud Apps or Defender for Endpoint standalone is readable on the learn pages; they point to Microsoft 365 E5 inclusion and separate pricing pages that returned 403 to a script.Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/microsoft-defender-endpoint), read 2026-11-16. - **microsoft-52** (vendor): Microsoft documentation on learn.microsoft.com is readable without sign-in, with last-updated dates (Agent 365 overview 2026-11-12).Quote: "Last updated on 2026-11-12"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/microsoft-agent-365/overview), read 2026-11-16. - **microsoft-53** (observed): Microsoft's MCP Gateway source is public on GitHub under the MIT license and offers telemetry, access control and observability hooks.Quote: "Enterprise-ready integration points including telemetry, access control and observability."Source: [github.com](https://github.com/microsoft/mcp-gateway), read 2026-11-16. - **microsoft-56** (third-party): On 2026-10-20, Help Net Security quoted Satya Nadella, CEO of Microsoft, on treating frontier models as insider risks, and said Project Perception shares signals with Agent 365.Quote: "Satya Nadella , CEO of Microsoft, wrote"Source: [helpnetsecurity.com](https://www.helpnetsecurity.com/2026/10/20/microsofts-ai-agents-threats-guardian-agents/), read 2026-11-16. - **microsoft-58** (none-found): None found on 2026-11-16: no named customer list or analyst page for shadow AI discovery, MCP security or AI runtime security was read on the Microsoft pages fetched.Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint), read 2026-11-16. - **microsoft-59** (vendor): Microsoft owns and sells these products itself: Defender for Endpoint is described as the endpoint security pillar of Microsoft Defender.Quote: "As the endpoint security pillar of Microsoft Defender , Defender for Endpoint feeds endpoint signals into the unified Defender portal."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/microsoft-defender-endpoint), read 2026-11-16. - **microsoft-61** (vendor): Microsoft Purview says DSPM for AI policies need at least 24 hours to collect data before results appear.Quote: "Allow at least 24 hours for these new policies to collect data to display the results in Data Security Posture Management for AI"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/purview/dspm-for-ai), read 2026-11-16. - **microsoft-64** (third-party): On 2026-10-19, SiliconANGLE covered the launch and named Satya Nadella as Microsoft Chairman and Chief Executive and Hayete Gallot as executive vice president of Microsoft Security.Quote: "Microsoft Chairman and Chief Executive Satya Nadella laid out the thinking behind the controls in an Oct. 10 essay"Source: [siliconangle.com](https://siliconangle.com/2026/10/19/microsoft-targets-the-agentic-insider-with-new-controls-for-rogue-ai-agents/), read 2026-11-16. - **microsoft-65** (third-party): On 2026-10-19, SiliconANGLE reported Microsoft's discovery tools can now spot 60 types of local agents on employee machines after 35 were added.Quote: "discovery tools can now spot 60 types of local agents on employee machines after 35 were added."Source: [siliconangle.com](https://siliconangle.com/2026/10/19/microsoft-targets-the-agentic-insider-with-new-controls-for-rogue-ai-agents/), read 2026-11-16. - **microsoft-68** (vendor): Microsoft says Agent 365 lets admins control which tools agents can connect to.Quote: "You control which tools agents can connect to."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/microsoft-agent-365/overview), read 2026-11-16. - **microsoft-69** (vendor): Microsoft says admins review and approve agents before publishing them to the organization and decide who can use them.Quote: "You review and approve agents before you publish them to your organization, and you decide who can use them."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/microsoft-agent-365/overview), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Lasso Security profile](/vendors/lasso-security/) - [Noma Security profile](/vendors/noma-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Noma Security MCP security: registry and tool controls URL: https://mcpservercybersecurity.com/vendors/noma-security/ Updated: 2026-11-16 Noma Security says it shows which MCP servers are installed, which agents use them and which tools they expose, including unapproved servers, with inventory gathered through existing EDR or MDM. It provides an MCP registry where servers are approved, reviewed or blocked, assesses supply chain risk, and inspects tool responses for hidden instructions, prompt injection and exfiltration patterns. Policy can be set by server, tool and user group, and enforcement runs through native hooks, gateways, SDKs and APIs. Third-party press confirmed Series A and B funding, and Noma publishes vulnerability research. Gaps on 16 Nov 2026: supported operating systems, deploy time, inspection of call arguments, SIEM export, public documentation and a price are none found. ## Strength and weakness for MCP Server Security **Strength:** Documents discovery, a registry with per-tool allow or block, and tool poisoning detection for MCP servers (noma-security-25, noma-security-27, noma-security-24). **Weakness:** No OS list or deploy time and no argument inspection stated; none found on 2026-11-16 (noma-security-29, noma-security-24). ## Scores and evidence Noma Security is 9th in the overall ranking at 51.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Noma Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 6 of 10 | 25 | The vendor claims inventory of servers, agents and tools via EDR or MDM and names Claude Code, Cursor, Copilot and others; vendor claims only. | [noma-security-08](#fact-noma-security-08), [noma-security-25](#fact-noma-security-25), [noma-security-09](#fact-noma-security-09), [noma-security-49](#fact-noma-security-49) | | Vetting before use | 6 of 10 | 20 | The vendor claims supply chain risk assessment and a registry where servers are approved or blocked; vendor claims only. | [noma-security-26](#fact-noma-security-26), [noma-security-27](#fact-noma-security-27), [noma-security-28](#fact-noma-security-28) | | Runtime control of tool calls | 5 of 10 | 20 | The vendor restricts tools and enforces via native hooks and inspects tool responses; inspection of call arguments: none found. | [noma-security-19](#fact-noma-security-19), [noma-security-23](#fact-noma-security-23), [noma-security-24](#fact-noma-security-24) | | Deployment | 4 of 10 | 15 | Discovery uses existing EDR or MDM; supported OS and a time to deploy: none found. | [noma-security-08](#fact-noma-security-08), [noma-security-34](#fact-noma-security-34), [noma-security-29](#fact-noma-security-29) | | Governance and audit | 5 of 10 | 5 | Policy by server, tool and user group, with observability and audit trails; SIEM export: none found. | [noma-security-37](#fact-noma-security-37), [noma-security-31](#fact-noma-security-31), [noma-security-48](#fact-noma-security-48), [noma-security-35](#fact-noma-security-35) | | Transparency | 3 of 10 | 10 | The vendor publishes vulnerability research; a price and public documentation: none found. | [noma-security-44](#fact-noma-security-44), [noma-security-45](#fact-noma-security-45), [noma-security-42](#fact-noma-security-42), [noma-security-43](#fact-noma-security-43) | | Proof | 6 of 10 | 5 | Third-party press confirms Series A and B funding; named customers and analyst coverage: none found. | [noma-security-03](#fact-noma-security-03), [noma-security-54](#fact-noma-security-54), [noma-security-05](#fact-noma-security-05), [noma-security-07](#fact-noma-security-07) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 9th | 51.5 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 9th | 52.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 9th | 51.5 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 8th | 54.0 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 9th | 51.5 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | joint 8th | 53.0 | ## Product, deployment, price and owner - Product assessed: Noma MCP Security (AI-SPM discovery plus Agent Access Control registry plus AI-DR tool poisoning detection) - Deployment: SaaS or on-prem; discovery via APIs, existing EDR or MDM; enforcement via hooks, gateways, SDKs and APIs - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 - Owner: Noma Security (standalone company; no parent stated on its pages) - Website: [noma.security](https://noma.security) ## Facts and sources - **noma-security-01** (vendor): Noma Security describes its offering as an agent security and governance platform for enterprise security teams.Quote: "The Agent Security & Governance Platform that helps enterprise security teams enable agent adoption without losing control."Source: [noma.security](https://noma.security/), read 2026-11-16. - **noma-security-02** (vendor): Noma says it is an AI security platform built for enterprise AI adoption; its about page lists Niv Braun as Co-Founder and CEO and Alon Tron as Co-Founder and CTO.Quote: "Noma is the AI security platform purpose-built to secure enterprise AI adoption."Source: [noma.security](https://noma.security/about), read 2026-11-16. - **noma-security-03** (third-party): TechCrunch reported on 2024-10-31 that Noma, founded in 2023, closed a $25 million Series A led by Ballistic Ventures, for $32 million raised in total, after a $7 million seed.Quote: "Noma raised $7 million in a previously undisclosed seed round led by Glilot Capital Partners with participation from Cyber Club London."Source: [techcrunch.com](https://techcrunch.com/2024/10/31/noma-is-building-tools-to-spot-security-issues-with-ai-apps/), read 2026-11-16. - **noma-security-04** (vendor): In a press release dated 2026-10-28, Noma lists its backers as Evolution Equity Partners, Ballistic Ventures, Glilot Capital, Cyber Club London, Databricks Ventures and SVCI.Quote: "Backed by Evolution Equity Partners, Ballistic Ventures, Glilot Capital, Cyber Club London, Databricks Ventures, and SVCI"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-named-to-fourth-annual-fortune-cyber-60-list-of-venture-backed-cybersecurity-companies-302922012.html), read 2026-11-16. - **noma-security-05** (vendor): Noma says on its homepage that over 10% of the Fortune 100 trusts it; no individual customer names were shown on the pages read.Quote: "Trusted by over 10% of the Fortune 100"Source: [noma.security](https://noma.security/), read 2026-11-16. - **noma-security-07** (vendor): On 2026-10-28 Noma announced it was named to the 2026-27 Fortune Cyber 60 list (early growth category), presented by Lightspeed Venture Partners and Amazon Web Services.Quote: "today announced it has been named to the 2026-27 Fortune Cyber 60, a list presented by Lightspeed Venture Partners and Amazon Web Services"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-named-to-fourth-annual-fortune-cyber-60-list-of-venture-backed-cybersecurity-companies-302922012.html), read 2026-11-16. - **noma-security-08** (vendor): Noma says it discovers endpoint agents, MCP servers and skills through the customer existing EDR or MDM, with no extra endpoint agent.Quote: "Noma discovers agents, MCP servers, and skills through your existing EDR or MDM. See where personal accounts and unapproved tools connect to company resources, without deploying another endpoint agent."Source: [noma.security](https://noma.security/platform/endpoint-agents), read 2026-11-16. - **noma-security-09** (vendor): On 2026-09-28 Noma announced endpoint agent security, saying enforcement coverage spans Claude Code, Claude Cowork, Cursor, Codex, Windsurf, Kiro, Antigravity and OpenClaw.Quote: "with coverage spanning agents including Claude Code, Claude Cowork, Cursor, Codex, Windsurf, Kiro, Antigravity, and OpenClaw"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-extends-agent-security-and-governance-to-the-employee-endpoint-302890925.html), read 2026-11-16. - **noma-security-18** (vendor): Noma says it enforces policy through AI gateways, MCP gateways, agent hooks, agent SDKs and direct APIs already in the customer environment.Quote: "Decouple governance from any single control point and enforce policy through AI gateways, MCP gateways, agent hooks, agent SDKs, and direct APIs, using the infrastructure already in your environment."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-19** (vendor): Noma says one policy set is enforced for agents on Bedrock behind an AI gateway, on Copilot Studio via integrations, and on developer machines via native hooks in Claude Code or Cursor.Quote: "The same behavioral monitoring and policy enforcement applies whether the agent runs on Bedrock behind an AI gateway"Source: [noma.security](https://noma.security/solutions/agent-behavior-runtime-control), read 2026-11-16. - **noma-security-23** (vendor): Noma says it restricts individual tools in approved MCP servers and enforces the decision when agents attempt to connect or act.Quote: "Restrict individual tools within approved servers and enforce those decisions when agents attempt to connect or act."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-24** (vendor): Noma says it inspects MCP tool responses for hidden instructions, prompt injection, workflow hijacking and data exfiltration patterns; inspection of call arguments is not stated.Quote: "Noma inspects MCP tool responses for hidden instructions, prompt injection, workflow hijacking, and data exfiltration patterns."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-25** (vendor): Noma says it shows which MCP servers are installed, which agents use them and which tools they expose, including unapproved servers.Quote: "See which MCP servers are installed, which agents use them, and which tools they expose."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-26** (vendor): Noma says it assesses MCP supply chain risk such as unpinned packages, excessive permissions and exposed secrets.Quote: "Find unpinned packages, excessive permissions, exposed secrets and more."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-27** (vendor): Noma says it provides an MCP registry where servers are approved, reviewed or blocked.Quote: "Approve, review, or block MCP servers."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-28** (vendor): Noma says AI-SPM inspects MCP servers, models and skills for risk and keeps a live registry of allowed AI components.Quote: "Inspect MCP servers, models, and skills for risk in the agent supply chain, and build a live registry of which AI components are allowed in the organization."Source: [noma.security](https://noma.security/products/ai-spm), read 2026-11-16. - **noma-security-29** (none-found): None found on 2026-11-16: supported operating systems or a time to deploy, on the endpoint agents and MCP security pages.Source: [noma.security](https://noma.security/platform/endpoint-agents), read 2026-11-16. - **noma-security-31** (vendor): Noma says AI-DR gives observability into every prompt, tool call, response and interaction across endpoint, SaaS and cloud agents.Quote: "complete observability into every prompt, tool call, response, and interaction your agents engage in across endpoint, SaaS, and cloud environments"Source: [noma.security](https://noma.security/products/ai-dr), read 2026-11-16. - **noma-security-34** (vendor): Noma says it supports both on-prem and SaaS deployments.Quote: "Support for both on-prem and SaaS deployments, so no model, training data, or security events leave your environment."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-35** (none-found): None found on 2026-11-16: named SIEM, SOAR, ticketing or EDR export integrations (beyond EDR/MDM discovery) on the AI-DR page.Source: [noma.security](https://noma.security/products/ai-dr), read 2026-11-16. - **noma-security-37** (vendor): Noma says policy can be scoped by server, individual tool and user group.Quote: "Noma scopes policy by server, by individual tool, and by user group"Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-42** (none-found): None found on 2026-11-16: a published price; noma.security/pricing returns Not Found and the sitemap lists no pricing page.Source: [noma.security](https://noma.security/pricing), read 2026-11-16. - **noma-security-43** (none-found): None found on 2026-11-16: public documentation; docs.noma.security redirects to a sign-in page.Source: [docs.noma.security](https://docs.noma.security), read 2026-11-16. - **noma-security-44** (vendor): Noma Labs published research on 2026-10-14 on Claude Code Desktop; the vendor says Anthropic scored the finding CVSS 8.7.Quote: "Anthropic scored Noma Labs' discovery as CVSS 8.7 (High) and treated the impact as equivalent to code execution on the host."Source: [noma.security](https://noma.security/noma-labs/sneakpeek-how-the-preview-pane-in-claude-code-desktop-can-exfiltrate-your-data), read 2026-11-16. - **noma-security-45** (vendor): Noma announced RufRoot, a CVSS 10.00 flaw its Labs found in the Ruflo agent platform (CVE-2026-59726); the release text read gives a 2026-06-30 publication date.Quote: "a maximum severity (CVSS 10.00) vulnerability discovered by Noma Labs in Ruflo"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-labs-discovers-critical-vulnerability-in-widely-adopted-open-source-ai-agent-platform-ruflo-302837494.html), read 2026-11-16. - **noma-security-46** (observed): Noma's public GitHub organization holds forks and a marketplace repo; no Noma-authored MCP security scanner was seen there.Quote: "Python SDK, Proxy Server (LLM Gateway) to call 100+ LLM APIs in OpenAI format"Source: [github.com](https://github.com/noma-security), read 2026-11-16. - **noma-security-48** (vendor): Noma says reporting includes inventory, framework mappings, test results and audit trails.Quote: "inventory, framework mappings, test results, and audit trails."Source: [noma.security](https://noma.security/products/ai-dr), read 2026-11-16. - **noma-security-49** (vendor): In its 2026-10-28 release Noma says it covers prebuilt agents like Claude Code, Cursor and GitHub Copilot running on developer machines.Quote: "prebuilt agents like Claude Code, Cursor, and GitHub Copilot running on developer machines"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-named-to-fourth-annual-fortune-cyber-60-list-of-venture-backed-cybersecurity-companies-302922012.html), read 2026-11-16. - **noma-security-54** (third-party): SiliconANGLE reported on 2025-07-31 that Noma raised a $100 million Series B led by Evolution Equity Partners with Ballistic Ventures and Glilot Capital.Quote: "Evolution Equity Partners led the Series B investment with participation from Ballistic Ventures and Glilot Capital."Source: [siliconangle.com](https://siliconangle.com/2025/07/31/ai-security-startup-noma-security-closes-100m-funding-round/), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Microsoft profile](/vendors/microsoft/) - [Pillar Security profile](/vendors/pillar-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Pillar Security MCP security: agentic endpoint URL: https://mcpservercybersecurity.com/vendors/pillar-security/ Updated: 2026-11-16 Pillar Security says its agentic endpoint product inventories local and remote MCP servers and scores each against a catalog of more than 23,000 pre-analyzed servers. The risk score covers reputation, supply chain, prompt handling, data access and actions, and per-server allow-lists restrict which tools agents may call in Monitor or Block mode, with decisions made in pre-tool hooks. Named clients include Claude Code, Cursor, Codex CLI and Gemini CLI, and deployment is by MDM script on macOS, Linux and Windows. Pillar has published research, which a third party confirmed. Limits on 16 Nov 2026: pre-action blocking is documented for Claude Code, Codex CLI and Cursor only, Claude Desktop and VS Code are not named, and a price and time to deploy are none found. ## Strength and weakness for MCP Server Security **Strength:** Inventories local and remote MCP servers, scores them against 23,000+ analyzed servers and enforces per-server tool allow-lists (pillar-security-13, pillar-security-14). **Weakness:** Pre-action blocking is documented for Claude Code, Codex CLI and Cursor only, and Claude Desktop and VS Code are not named (pillar-security-05, pillar-security-09). ## Scores and evidence Pillar Security is 6th in the overall ranking at 57.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Pillar Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 6 of 10 | 25 | The vendor names Claude Code, Cursor, Codex CLI, Gemini CLI and others, and inventories local and remote MCP servers; Claude Desktop and VS Code are not named, and all is vendor-stated. | [pillar-security-04](#fact-pillar-security-04), [pillar-security-05](#fact-pillar-security-05), [pillar-security-08](#fact-pillar-security-08), [pillar-security-13](#fact-pillar-security-13) | | Vetting before use | 6 of 10 | 20 | The vendor scores servers against a catalog on reputation, supply chain and actions, with per-server tool allowlists; vendor claims only. | [pillar-security-13](#fact-pillar-security-13), [pillar-security-14](#fact-pillar-security-14) | | Runtime control of tool calls | 6 of 10 | 20 | The vendor decides on every action in pre-tool hooks with Monitor or Block and tool allowlists; argument inspection detail: none found. | [pillar-security-09](#fact-pillar-security-09), [pillar-security-14](#fact-pillar-security-14), [pillar-security-12](#fact-pillar-security-12) | | Deployment | 6 of 10 | 15 | MDM scripts on macOS, Linux and Windows; a time to deploy: none found. | [pillar-security-06](#fact-pillar-security-06), [pillar-security-15](#fact-pillar-security-15), [pillar-security-33](#fact-pillar-security-33) | | Governance and audit | 4 of 10 | 5 | Controls by mode and an activity view are documented; SIEM export for MCP tool-call logs and per-team policy: none found. | [pillar-security-12](#fact-pillar-security-12), [pillar-security-48](#fact-pillar-security-48) | | Transparency | 4 of 10 | 10 | Research is published and confirmed by a third party; a price and public documentation: none found. | [pillar-security-36](#fact-pillar-security-36), [pillar-security-42](#fact-pillar-security-42), [pillar-security-34](#fact-pillar-security-34), [pillar-security-35](#fact-pillar-security-35) | | Proof | 6 of 10 | 5 | Third-party reports confirm seed funding and a Fortune 300 trial, with named testimonials; analyst recognition is a vendor claim. | [pillar-security-39](#fact-pillar-security-39), [pillar-security-41](#fact-pillar-security-41), [pillar-security-37](#fact-pillar-security-37), [pillar-security-40](#fact-pillar-security-40) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 6th | 57.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 6th | 58.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 6th | 58.0 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 7th | 57.0 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 7th | 57.0 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | joint 8th | 53.0 | ## Product, deployment, price and owner - Product assessed: Agentic Endpoint (MCP inventory, risk scores, tool allow-lists) - Deployment: Managed cloud, self-hosted or hybrid; endpoint product by MDM script, with a per-user service and agent hooks - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 - Owner: Pillar Security. Vanta announced a definitive agreement to acquire it on 9 Nov 2026; the deal was not stated as closed on 16 Nov 2026 - Website: [pillar.security](https://www.pillar.security) ## Facts and sources - **pillar-security-01** (vendor): On 2026-11-09 Pillar's founders announced that Vanta has signed a definitive agreement to acquire Pillar Security, so the deal is announced and not described as closed.Quote: "Today, Vanta has signed a definitive agreement to acquire Pillar Security."Source: [pillar.security](https://www.pillar.security/blog/vanta-to-acquire-pillar-security-building-trust-in-ai-together), read 2026-11-16. - **pillar-security-02** (vendor): Pillar says it was founded in 2023 and describes its product as discovering, testing and securing AI applications and agents and blocking unauthorized actions at runtime.Quote: "When we founded Pillar in 2023, most of the conversation around AI security was about chatbots and prompts."Source: [pillar.security](https://www.pillar.security/blog/vanta-to-acquire-pillar-security-building-trust-in-ai-together), read 2026-11-16. - **pillar-security-03** (vendor): Pillar says its platform discovers, tests, secures and governs AI agents, across five modules: AI Discovery & Posture, Red Teaming & Attack Surface Exposure, Runtime Guardrails, Agentic Endpoint, and Governance & Compliance.Quote: "One platform to discover, test, secure and govern every AI agent across your organization"Source: [pillar.security](https://www.pillar.security), read 2026-11-16. - **pillar-security-04** (vendor): Pillar says its Agentic Endpoint product finds coding agents, MCP servers, skills and plugins on every managed machine.Quote: "Find coding agents, MCP servers, skills and plugins on every machine."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-05** (vendor): Pillar names the agents it scans in depth (Claude Code, Cursor, Codex CLI, Gemini CLI, OpenCode, Pi) and others it detects by fingerprint (Amazon Q, Antigravity, OpenClaw).Quote: "Deep scans for Claude Code, Cursor, Codex CLI, Gemini CLI, OpenCode and Pi; fingerprinting for Amazon Q, Antigravity, OpenClaw and more."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-06** (vendor): Pillar says the endpoint product is deployed through MDM scripts (Jamf Pro, Kandji, JumpCloud or any MDM that runs scripts) on macOS, Linux and Windows.Quote: "Roll out with Jamf Pro, Kandji, JumpCloud or any MDM that runs scripts, on macOS, Linux and Windows."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-08** (vendor): Pillar says it audits agent configuration such as permission rules, approval and sandbox modes, MCP servers, hooks and plugins, and maps issues to its SAIL framework.Quote: "Pillar parses each agent's permission rules, approval and sandbox modes, MCP servers, hooks and plugins"Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-09** (vendor): Pillar says Runtime Guardian sits in the pre-tool hook of Claude Code, Codex CLI and Cursor and decides on every action before it runs.Quote: "Runtime Guardian sits in the agent's own pre-tool hook for Claude Code, Codex CLI and Cursor, and decides on every action before it runs."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-12** (vendor): Pillar says each control can be set to Off, Monitor or Block, and every blocked or asked decision appears in an Activity view with the full session timeline.Quote: "Every decision, blocked or asked, appears in Activity with the full session timeline."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-13** (vendor): Pillar says it inventories local and remote MCP servers and scores each against a catalog of more than 23,000 pre-analyzed servers.Quote: "Risk-score every MCP server against 23,000+ pre-analyzed servers."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-14** (vendor): Pillar says each MCP risk score covers reputation, supply chain, prompt handling, data access and actions, and per-server allow-lists restrict which tools agents may call, in Monitor or Block mode.Quote: "Restrict which tools each MCP server may invoke, in Monitor or Block mode."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-15** (vendor): Pillar says posture scans read agent configuration and not source code, and that secrets in configuration are removed before anything leaves the machine.Quote: "Secrets in agent configuration are removed before anything leaves the machine."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-29** (vendor): Pillar says it runs as managed cloud, self-hosted in the customer's cloud account or data center, or hybrid.Quote: "Deploy Pillar where your data lives: our managed cloud, your own cloud account or data center, or a mix of both."Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-30** (vendor): Pillar's go-live steps are scope, install, connect (AI gateway, code repositories, device management) and go live in monitor mode before blocking.Quote: "Start in monitor mode, then switch to blocking."Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-33** (none-found): None found on 2026-11-16: a published time to deploy or time to first inventory, searched on the deployment, agentic endpoint and discovery pages.Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-34** (none-found): None found on 2026-11-16: a published price; the pricing URL returned 404 and the site offers only a demo request.Source: [pillar.security](https://www.pillar.security/pricing), read 2026-11-16. - **pillar-security-35** (none-found): None found on 2026-11-16: public product documentation; docs.pillar.security loads a sign-in page marked noindex.Source: [docs.pillar.security](https://docs.pillar.security), read 2026-11-16. - **pillar-security-36** (vendor): Pillar publishes the SAIL framework and a research blog and says SAIL has passed 50,000 downloads and maps 91 AI risks across seven lifecycle phases.Quote: "SAIL has passed 50,000 downloads, and maps 91 AI risks across seven lifecycle phases."Source: [pillar.security](https://www.pillar.security/blog/vanta-to-acquire-pillar-security-building-trust-in-ai-together), read 2026-11-16. - **pillar-security-37** (vendor): Pillar's site shows testimonials from named people at Eleos, another company and Similarweb.The exact quote is on the source page.Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-38** (vendor): Pillar says a Fortune 300 manufacturer runs it fully self-hosted.Quote: "A Fortune 300 manufacturer runs Pillar fully self-hosted."Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-39** (third-party): CRN reported that Pillar announced a $9 million seed round led by Shield Capital in April 2025, and listed it among its 10 hottest AI security startups of 2026.Quote: "Pillar announced its $9 million seed funding round led by Shield Capital in April 2025."Source: [crn.com](https://www.crn.com/news/security/2026/the-10-hottest-ai-security-startups-of-2026-so-far?page=6), read 2026-11-16. - **pillar-security-40** (vendor): The Pillar newsroom lists a Gartner Cool Vendor in AI Software Security recognition dated Jul 1, 2026 and a Frost & Sullivan award dated Sep 17, 2025.Quote: "Pillar Security Named as a 2026 Gartner® Cool Vendor in AI Software Security"Source: [pillar.security](https://www.pillar.security/newsroom), read 2026-11-16. - **pillar-security-41** (third-party): A Refresh Miami profile dated 2026-01-26 says Pillar scanned more than 30,000 repositories in a Fortune 300 proof of concept and found 8,000 AI models and 400 AI systems unknown to the security team.Quote: "Pillar scanned more than 30,000 code repositories. The result: 8,000 AI models and 400 AI systems"Source: [refreshmiami.com](https://refreshmiami.com/news/ai-broke-the-security-model-pillar-wants-to-fix-it/), read 2026-11-16. - **pillar-security-42** (third-party): BleepingComputer reported on 2026-07-20 that Pillar researchers published a Week of Sandbox Escapes series on sandbox bypasses in AI coding agents, with findings sorted into four failure modes.Quote: "Pillar sorts the seven findings into four failure modes:"Source: [bleepingcomputer.com](https://www.bleepingcomputer.com/news/security/cursor-codex-gemini-cli-antigravity-hit-by-sandbox-escapes/), read 2026-11-16. - **pillar-security-48** (none-found): None found on 2026-11-16: a published exportable SIEM format for MCP tool-call logs; SIEM streaming is stated only for runtime guardrails and the self-hosted integrations list.Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Noma Security profile](/vendors/noma-security/) - [SentinelOne (Prompt Security) profile](/vendors/sentinelone/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # SentinelOne MCP security: Prompt Security URL: https://mcpservercybersecurity.com/vendors/sentinelone/ Updated: 2026-11-16 SentinelOne says Prompt Security discovers every AI agent and MCP server in the environment, including the connectors, tools, skills and plugins an agent touches, and scores each agent and MCP server for risk against the customer's policy. It names Copilot, Cursor and Claude Code, offers a browser extension, a dedicated AI endpoint agent, an MCP gateway and a self-hosted on-premises option, and says setup takes minutes. SentinelOne says the platform produces a searchable log of every session, action and tool an agent touches. Third-party press confirmed funding and the acquisition. Gaps on 16 Nov 2026: named MCP client configurations, an approved-server registry, argument inspection, supported operating systems, SIEM export and a product price are none found. ## Strength and weakness for MCP Server Security **Strength:** Documents discovery of MCP servers, risk scoring and a searchable action log (sentinelone-10, sentinelone-45, sentinelone-44). **Weakness:** None found on 2026-11-16 for named MCP clients, an approved server registry or argument inspection (sentinelone-49). ## Scores and evidence SentinelOne (Prompt Security) is 14th in the overall ranking at 36.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **SentinelOne (Prompt Security): score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 4 of 10 | 25 | The vendor names Copilot, Cursor and Claude Code and claims discovery of MCP servers and their tools; named MCP client configurations on the agentic page: none found. | [sentinelone-09](#fact-sentinelone-09), [sentinelone-10](#fact-sentinelone-10), [sentinelone-49](#fact-sentinelone-49) | | Vetting before use | 3 of 10 | 20 | The vendor claims risk scoring of every MCP server; an enforced approved list or detail on what is scanned: none found. | [sentinelone-45](#fact-sentinelone-45), [sentinelone-56](#fact-sentinelone-56), [sentinelone-49](#fact-sentinelone-49) | | Runtime control of tool calls | 3 of 10 | 20 | Runtime claims are general (stop policy violations, endpoint enforcement); inspection of tool-call arguments: none found. | [sentinelone-32](#fact-sentinelone-32), [sentinelone-43](#fact-sentinelone-43), [sentinelone-40](#fact-sentinelone-40), [sentinelone-49](#fact-sentinelone-49) | | Deployment | 4 of 10 | 15 | The vendor names a browser extension and a dedicated endpoint agent and claims setup in minutes; supported operating systems and a measured deploy time: none found. | [sentinelone-13](#fact-sentinelone-13), [sentinelone-14](#fact-sentinelone-14), [sentinelone-55](#fact-sentinelone-55), [sentinelone-18](#fact-sentinelone-18) | | Governance and audit | 3 of 10 | 5 | The vendor claims a searchable log of agent sessions and tools; SIEM export or per-team policy: none found. | [sentinelone-44](#fact-sentinelone-44), [sentinelone-45](#fact-sentinelone-45), [sentinelone-48](#fact-sentinelone-48) | | Transparency | 4 of 10 | 10 | The vendor publishes research and an observed open-source skill suite; a product price and public documentation: none found. | [sentinelone-27](#fact-sentinelone-27), [sentinelone-28](#fact-sentinelone-28), [sentinelone-25](#fact-sentinelone-25), [sentinelone-26](#fact-sentinelone-26) | | Proof | 6 of 10 | 5 | Third-party press confirms funding and the acquisition, and named CISOs are quoted; independent analyst coverage: none found. | [sentinelone-03](#fact-sentinelone-03), [sentinelone-04](#fact-sentinelone-04), [sentinelone-05](#fact-sentinelone-05), [sentinelone-29](#fact-sentinelone-29) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 14th | 36.5 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 14th | 37.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 14th | 39.0 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 14th | 34.5 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 13th | 34.5 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 13th | 35.0 | ## Product, deployment, price and owner - Product assessed: Prompt Security Agentic AI Security, MCP discovery and risk scoring (SentinelOne) - Deployment: Singularity Platform console; browser extension, dedicated AI endpoint agent, MCP Gateway and API integrations; self-hosted Prompt Security On-Premises; AgentCore gateway integration - Time to deploy: SentinelOne says it deploys in minutes; no measured figure or OS list found on 2026-11-16 - Price: none found on 2026-11-16 for Prompt Security; the pricing page lists Singularity tiers and says Enterprise is Contact Sales for Pricing - Owner: SentinelOne, Inc. (NYSE: S), Mountain View, California - Website: sentinelone.com ## Facts and sources - **sentinelone-01** (vendor): SentinelOne says Prompt Security is built into its Singularity Platform and managed from the same console as its endpoint, cloud, identity and data security.Quote: "Prompt Security is built into the SentinelOne Singularity Platform."Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-02** (vendor): On 2025-08-05 SentinelOne announced a definitive agreement to acquire Prompt Security, so the AI security line sits inside SentinelOne (NYSE: S).Quote: "signed a definitive agreement to acquire Prompt Security"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-to-acquire-prompt-security-to-advance-genai-security/), read 2026-11-16. - **sentinelone-03** (third-party): CRN reports that Prompt Security was founded in 2023 and raised an $18 million Series A in November 2024 before SentinelOne agreed to acquire it.Quote: "Founded in 2023, Prompt Security offers protection against AI-related threats"Source: [crn.com](https://www.crn.com/news/security/2025/sentinelone-to-acquire-ai-protection-startup-prompt-security/), read 2026-11-16. - **sentinelone-04** (third-party): CRN reports the Series A was led by Jump Capital with Hetz Ventures, Ridge Ventures, Okta and F5 taking part.Quote: "In November 2024, Prompt Security announced an $18 million Series A funding round led by Jump Capital"Source: [crn.com](https://www.crn.com/news/security/2025/sentinelone-to-acquire-ai-protection-startup-prompt-security/), read 2026-11-16. - **sentinelone-05** (third-party): Techzine reports SentinelOne acquired Israeli company Prompt Security to prepare its Singularity platform for secure AI use in organizations.Quote: "SentinelOne has announced the acquisition of Israeli company Prompt Security."Source: [techzine.eu](https://www.techzine.eu/news/security/133547/sentinelone-strengthens-ai-protection-with-prompt-security-deal/), read 2026-11-16. - **sentinelone-06** (third-party): Wikipedia lists SentinelOne as a public company headquartered in Mountain View, California, founded in 2013, with Tomer Weingarten as chairman and CEO.Quote: "Headquarters Mountain View, California , U.S."Source: [en.wikipedia.org](https://en.wikipedia.org/wiki/SentinelOne), read 2026-11-16. - **sentinelone-09** (vendor): SentinelOne says Prompt Security tracks AI code assistants in the environment and names GitHub Copilot, Cursor and Claude Code.Quote: "Track every AI code assistant in the environment, including GitHub Copilot, Cursor, and Claude Code"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-ai-usage-control/), read 2026-11-16. - **sentinelone-10** (vendor): SentinelOne says Prompt Security discovers every AI agent and MCP server in the environment, including the connectors, tools, skills and plugins an agent touches.Quote: "See every connector, tool, skill, and plugin an agent touches, and who's running it"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-13** (vendor): Prompt Security says AI governance needs local sensors, namely a browser extension and a dedicated AI endpoint agent, because network controls alone fall short.Quote: "you need local, purpose-built sensors: a browser extension and a dedicated AI endpoint agent"Source: [prompt.security](https://prompt.security/blog/the-key-layer-in-ai-security-browser-and-endpoint-sensors), read 2026-11-16. - **sentinelone-14** (vendor): SentinelOne says an MCP Gateway, browser extension, security plugins and API integrations work together so no agent surface goes undetected.Quote: "MCP Gateway, browser extension, security plugins, and API integrations work together so no agent surface goes undetected"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-17** (vendor): SentinelOne says Prompt Security deploys in minutes with no complex integrations or infrastructure changes.Quote: "Prompt Security deploys in minutes. No complex integrations or infrastructure changes are required."Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-18** (none-found): None found on 2026-11-16: looked on the Prompt Security product pages for supported operating systems, browsers, a measured time to first AI inventory and a published deployment requirements list; only "minutes" was found.Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-25** (vendor): SentinelOne's public pricing page lists tiers but says the Enterprise tier is "Contact Sales for Pricing" and does not list Prompt Security.Quote: "Contact Sales for Pricing"Source: [sentinelone.com](https://www.sentinelone.com/pricing/), read 2026-11-16. - **sentinelone-26** (none-found): None found on 2026-11-16: looked on the pricing page and Prompt Security pages for a published price of Prompt Security or AI Usage Control and for public technical documentation readable without sign-in; found marketing pages and a demo form only.Source: [sentinelone.com](https://www.sentinelone.com/pricing/), read 2026-11-16. - **sentinelone-27** (vendor): On 2026-08-20 Prompt Security published research on three agent cases (AGENTS.md goal hijack, marketplace skills dependency hijack, instruction file integrity) and says approval prompts should show the real destination and payload.Quote: "approval prompts that show the real destination and payload instead of a friendly summary"Source: [prompt.security](https://prompt.security/blog/when-ai-agents-become-the-supply-chain-hidden-control-planes-in-agentic-systems), read 2026-11-16. - **sentinelone-28** (observed): Prompt Security publishes ClawSec on GitHub under AGPL-3.0, a security skill suite for OpenClaw, Hermes, PicoClaw and NanoClaw agents with 1.1k stars.Quote: "ClawSec source code is licensed under GNU AGPL-3.0-or-later"Source: [github.com](https://github.com/prompt-security/clawsec), read 2026-11-16. - **sentinelone-29** (vendor): SentinelOne's page quotes Margarita Rivera, Global CISO of Carnival Corporation, on Prompt Security giving a first layer of visibility and protection.Quote: "Prompt Security has been incredible for our organization."Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-ai-usage-control/), read 2026-11-16. - **sentinelone-30** (vendor): SentinelOne's page quotes Richard Moore, CISO at 10x Banking, on Prompt Security.Quote: "Prompt Security has allowed us to be enablers for the business."Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-ai-usage-control/), read 2026-11-16. - **sentinelone-32** (vendor): Prompt Security says a dedicated AI endpoint agent should detect new agents, inspect prompt construction, enforce runtime policies and prevent sensitive data exposure before execution.Quote: "Prevent sensitive data exposure before execution"Source: [prompt.security](https://prompt.security/blog/the-key-layer-in-ai-security-browser-and-endpoint-sensors), read 2026-11-16. - **sentinelone-33** (vendor): Prompt Security says monitoring and controlling local MCP communications requires endpoint-level visibility and enforcement.Quote: "Monitoring and controlling MCP communications requires endpoint-level visibility and enforcement."Source: [prompt.security](https://prompt.security/blog/the-key-layer-in-ai-security-browser-and-endpoint-sensors), read 2026-11-16. - **sentinelone-34** (vendor): On 2026-06-17 SentinelOne announced an upcoming Prompt Security integration with Amazon Bedrock AgentCore supplying signals for prompt injection, PII exposure, tool-use validation, LLM response monitoring and data leakage prevention.Quote: "prompt injection detection, PII exposure, tool-use validation, LLM response monitoring, and data leakage prevention"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-announces-ai-security-integration-with-amazon-bedrock-agentcore-to-deliver-runtime-guardrails-for-ai-agents/), read 2026-11-16. - **sentinelone-40** (third-party): Techzine reports Prompt Security works standalone of the AI model and via the Model Context Protocol through MCP gateway security.Quote: "also works via the now established Model Context Protocol via MCP gateway security"Source: [techzine.eu](https://www.techzine.eu/news/security/133547/sentinelone-strengthens-ai-protection-with-prompt-security-deal/), read 2026-11-16. - **sentinelone-43** (vendor): SentinelOne says Prompt Security stops agent policy violations before they propagate across connected systems.Quote: "Stop violations before they propagate across connected systems"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-44** (vendor): SentinelOne says Prompt Security produces a searchable log of every session, action and tool an agent touches.Quote: "Produce a searchable log of every session, action, and tool an agent touches."Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-45** (vendor): SentinelOne says Prompt Security scores every agent and MCP server for risk against the customer's policy.Quote: "Score every agent and MCP server for risk against your policy"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-46** (vendor): On 2026-03-23 SentinelOne said Prompt Security On-Premises is self-hosted, works in disconnected environments, and blocks prompt injections and data leaks as a firewall for internal and external AI tools.Quote: "Prompt Security blocks threats like prompt injections and data leaks"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-brings-ai-security-to-on-premise-regulated-sovereign-self-hosted-and-airgapped-environments/), read 2026-11-16. - **sentinelone-48** (none-found): None found on 2026-11-16: looked on the Prompt Security pages and press releases for a named SIEM export (beyond Singularity AI SIEM), SOAR, EDR, identity provider or ticketing integration for Prompt Security events; none named.Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-49** (none-found): None found on 2026-11-16: looked on the Agentic AI Security page for named MCP clients (Claude Desktop, Cursor, VS Code, Windsurf, Codex CLI), an approved server registry or allowlist, and inspection of tool-call arguments; none named.Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-52** (vendor): Prompt Security says a skill can redirect a package install to a substituted source, and recommends approved sources only and pinned versions for what agents may install.Quote: "approved sources only, pinned versions, no shortcuts through unofficial channels"Source: [prompt.security](https://prompt.security/blog/when-ai-agents-become-the-supply-chain-hidden-control-planes-in-agentic-systems), read 2026-11-16. - **sentinelone-55** (vendor): SentinelOne says setup is to connect your environment and integrate with existing infrastructure in minutes, with protection starting immediately.Quote: "Integrate with your existing infrastructure in minutes. No disruption. Protection starts immediately."Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-56** (vendor): SentinelOne's Agentic AI Security page cites 15,000+ MCP servers assessed for risk across the environment.Quote: "MCP servers assessed for risk across your environment"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Pillar Security profile](/vendors/pillar-security/) - [Snyk (Agent Scan, Evo ADS) profile](/vendors/snyk/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Snyk MCP security: Agent Scan and Evo ADS URL: https://mcpservercybersecurity.com/vendors/snyk/ Updated: 2026-11-16 Snyk's Agent Scan is an Apache-2.0 command-line scanner whose README lists the clients it reads on each operating system, including Claude Desktop, Claude Code, Cursor and VS Code, and its fleet view reports servers and tools. Evo ADS adds vetting: a local stdio server is matched against the Snyk catalog and assessed from its signature without starting it, using trust signals such as popularity, first-party status, URL reputation and maintainer identity. Snyk says Evo can block MCP server execution on the developer's machine at invocation. It publishes a rate of 1.0 credit per active machine per day, and Evo needs an Enterprise subscription. Limits: blocking covers Claude Code, Codex, Cursor and GitHub Copilot, audit data is kept seven days, and argument inspection and SIEM export: none found on 16 Nov 2026. ## Strength and weakness for MCP Server Security **Strength:** Covers all three jobs for developer machines in one product: discovery across many IDEs and CLI agents (snyk-10, snyk-11), catalog-based vetting with risk scores (snyk-14, snyk-15) and blocking of non-allowed MCP servers at invocation on the machine (snyk-19, snyk-24), with a published per-machine price (snyk-35) and Apache-2.0 open-source scanner (snyk-07). **Weakness:** Runtime blocking covers only Claude Code, Codex, Cursor and GitHub Copilot (snyk-21), and none found on 2026-11-16 for tool-call argument inspection (snyk-23) or SIEM export and per-team policy (snyk-28); audit data is kept seven days (snyk-27) and Evo needs an Enterprise subscription (snyk-36). ## Scores and evidence Snyk (Agent Scan, Evo ADS) is 1st in the overall ranking at 77.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Snyk (Agent Scan, Evo ADS): score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 8 of 10 | 25 | The open-source scanner, observed on GitHub, documents a per-OS table of many clients including Claude Desktop, Claude Code, Cursor and VS Code, and a fleet view of servers and tools. | [snyk-09](#fact-snyk-09), [snyk-11](#fact-snyk-11), [snyk-12](#fact-snyk-12), [snyk-13](#fact-snyk-13), [snyk-06](#fact-snyk-06) | | Vetting before use | 9 of 10 | 20 | The vendor documents catalog matching, trust signals and tool-description risk indexes, with an observed open-source scanner and a third-party report; uncataloged stdio servers get no assessment. | [snyk-14](#fact-snyk-14), [snyk-15](#fact-snyk-15), [snyk-16](#fact-snyk-16), [snyk-24](#fact-snyk-24), [snyk-06](#fact-snyk-06), [snyk-41](#fact-snyk-41) | | Runtime control of tool calls | 6 of 10 | 20 | The vendor blocks unauthorized MCP server calls on the workstation, confirmed in third-party press; inspection of tool-call arguments: none found. | [snyk-19](#fact-snyk-19), [snyk-20](#fact-snyk-20), [snyk-22](#fact-snyk-22), [snyk-41](#fact-snyk-41), [snyk-23](#fact-snyk-23) | | Deployment | 8 of 10 | 15 | MDM install via Jamf or Intune with OS minimums and an observed PyPI command-line package; a fleet time to deploy: none found. | [snyk-29](#fact-snyk-29), [snyk-30](#fact-snyk-30), [snyk-33](#fact-snyk-33), [snyk-08](#fact-snyk-08), [snyk-31](#fact-snyk-31) | | Governance and audit | 4 of 10 | 5 | Allowlists and risk-based policies with audit logging are documented, with seven-day retention; SIEM export and team scoping: none found. | [snyk-24](#fact-snyk-24), [snyk-25](#fact-snyk-25), [snyk-26](#fact-snyk-26), [snyk-27](#fact-snyk-27), [snyk-28](#fact-snyk-28) | | Transparency | 9 of 10 | 10 | Documentation is public, a per-machine price is published (enterprise subscription required), and the scanner is observed as Apache-2.0 open source. | [snyk-37](#fact-snyk-37), [snyk-35](#fact-snyk-35), [snyk-06](#fact-snyk-06), [snyk-07](#fact-snyk-07) | | Proof | 8 of 10 | 5 | Third-party press confirms large funding, a named customer is quoted, and an analyst Leader placement is a vendor claim. | [snyk-42](#fact-snyk-42), [snyk-39](#fact-snyk-39), [snyk-40](#fact-snyk-40) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 1st | 77.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 1st | 76.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 1st | 78.5 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | 1st | 77.5 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | 2nd | 70.0 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 3rd | 67.0 | ## Product, deployment, price and owner - Product assessed: Evo Agentic Development Security (Agent Supply Chain Security, Agent Behavior Governance with MCP Governance) and the open-source Snyk Agent Scan (formerly Invariant Labs mcp-scan); Snyk Code, Open Source and other AppSec engines not counted - Deployment: Installer deployed with MDM (Jamf Pro on macOS, Microsoft Intune on Windows) or by command line on one machine, binding to a Snyk tenant with a push key; scans run every 24 hours; the open-source CLI runs with uvx or a standalone binary - Time to deploy: First install run can take several minutes (snyk-31); no fleet time-to-deploy figure published, none found on 2026-11-16 - Price: Rate card on snyk.io/plans: Evo ADS Coding Agent Security is 1.0 credit per active machine per day, 1 credit = $1; requires an Enterprise Platform Subscription (contact sales); Agent Scan CLI is open source - Owner: Snyk Limited - Website: snyk.io ## Facts and sources - **snyk-01** (vendor): Snyk's own About page says the company was founded in 2015 and is headquartered in Boston.Quote: "Founded in 2015 and recognized with unicorn status in 2020"Source: [snyk.io](https://snyk.io/about/), read 2026-11-16. - **snyk-02** (vendor): Snyk says it started in London and Tel Aviv and is headquartered in Boston.Quote: "We started in London and Tel Aviv, are headquartered in Boston"Source: [snyk.io](https://snyk.io/about/), read 2026-11-16. - **snyk-03** (vendor): Snyk's leadership page lists Manoj Nair as Chief Innovation Officer and CTO, leading the office responsible for incubation and acquisition strategy; the same page lists Ken MacAskill as CEO and CFO.Quote: "Manoj leads Snyk's Emerging Technologies and Solutions Office (ETSO)."Source: [snyk.io](https://snyk.io/about/leadership/), read 2026-11-16. - **snyk-04** (vendor): On 2025-06-24 Snyk announced it acquired Invariant Labs, the research firm that built agent guardrails and MCP server scanning.Quote: "With Invariant Labs, developers can inspect and observe agent behavior, enforce contextual security rules on agent systems, and scan MCP servers for vulnerabilities."Source: [snyk.io](https://snyk.io/news/snyk-acquires-invariant-labs-to-accelerate-agentic-ai-security-innovation/), read 2026-11-16. - **snyk-05** (observed): The former Invariant Labs repository invariantlabs-ai/mcp-scan now resolves to snyk/agent-scan on GitHub, owned by the snyk organization, so mcp-scan is a Snyk product today.Quote: "full_name: snyk/agent-scan, owner login: snyk"Source: [api.github.com](https://api.github.com/repos/invariantlabs-ai/mcp-scan), read 2026-11-16. - **snyk-06** (observed): GitHub describes snyk/agent-scan as an Apache-2.0 licensed security scanner for AI agents, MCP servers and agent skills, with 3,242 stars and a push on 2026-11-16.Quote: "Security scanner for AI agents, MCP servers and agent skills."Source: [api.github.com](https://api.github.com/repos/snyk/agent-scan), read 2026-11-16. - **snyk-07** (observed): The agent-scan source code is open source under the Apache License 2.0 according to the GitHub API record.Quote: ""name": "Apache License 2.0""Source: [api.github.com](https://api.github.com/repos/snyk/agent-scan), read 2026-11-16. - **snyk-08** (observed): PyPI lists snyk-agent-scan version 0.6.11 under the Apache-2.0 license as an agent supply chain security scanner.Quote: "Agent supply chain security scanner."Source: [pypi.org](https://pypi.org/pypi/snyk-agent-scan/json), read 2026-11-16. - **snyk-09** (vendor): Snyk's Agent Scan README says it auto-discovers MCP configurations, agent tools and skills on the machine it runs on.Quote: "Auto-discover MCP configurations, agent tools, skills"Source: [raw.githubusercontent.com](https://raw.githubusercontent.com/snyk/agent-scan/main/README.md), read 2026-11-16. - **snyk-10** (vendor): The README names the clients it scans: Claude, Cursor, GitHub Copilot, Windsurf, Gemini CLI, Amp, Amazon Q and others.Quote: "Scanning of Claude, Cursor, GitHub Copilot, Windsurf, Gemini CLI, Amp, Amazon Q, and other agents."Source: [raw.githubusercontent.com](https://raw.githubusercontent.com/snyk/agent-scan/main/README.md), read 2026-11-16. - **snyk-11** (vendor): The README has a per-OS table covering Windsurf, Cursor, VS Code, GitHub Copilot, Claude Desktop, Claude Code, Gemini CLI, Kiro, OpenCode, Antigravity, Codex, Amp, OpenClaw and Amazon Q, and it notes Copilot in VS Code is detected via the shared Copilot home.Quote: "GitHub Copilot covers the CLI and the desktop app, which share the `~/.copilot` home with Copilot in VS Code"Source: [raw.githubusercontent.com](https://raw.githubusercontent.com/snyk/agent-scan/main/README.md), read 2026-11-16. - **snyk-12** (vendor): Snyk's Evo docs say the fleet view lists every machine that reported a scan, with counts of MCP servers and skills discovered and scanned.Quote: "Your fleet shows every machine that has reported a scan, including how many MCP servers and skills Evo discovered and how many it successfully scanned."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md), read 2026-11-16. - **snyk-13** (vendor): Snyk's docs say that for remote MCP servers the scanner contacts the server directly to retrieve its current tools and capabilities.Quote: "Agent Supply Chain Security contacts a remote server directly to retrieve its current tools and capabilities."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md), read 2026-11-16. - **snyk-14** (vendor): Snyk's docs say a local stdio MCP server is matched against the Snyk catalog and assessed from its signature without starting it.Quote: "Agent Supply Chain Security matches a local (stdio) MCP server against the Snyk catalog and assesses it from that signature. It does not start the server."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md), read 2026-11-16. - **snyk-15** (vendor): Snyk's MCP Risk Score includes trust signals: server popularity, official first-party status, URL reputation and maintainer identity.Quote: "how popular the server is, whether it is an official first-party MCP server, the reputation of the URLs it connects to, and the identity of its maintainers."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md), read 2026-11-16. - **snyk-16** (vendor): Snyk's docs list MCP tool-description risk indexes, including prompt injection in a tool, untrusted content, private data and destructive capabilities, each scored 0 to 1,000.Quote: "Prompt injection in a tool: an agent processes hidden instructions within a tool description as commands."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md), read 2026-11-16. - **snyk-17** (vendor): Snyk's docs admit a limit: a stdio server not yet in the Snyk catalog ends as Discovery only, with no risk assessment.Quote: "For stdio MCP servers, a scan also ends in Discovery only when the server is not yet in Snyk's catalog of regularly assessed servers."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md), read 2026-11-16. - **snyk-18** (vendor): The open-source Agent Scan says that to retrieve tool descriptions it starts stdio MCP servers by running the commands in the config, and advises sandboxing and consent prompts.Quote: "To retrieve tool descriptions, Agent Scan starts stdio MCP servers by executing the commands in the config"Source: [raw.githubusercontent.com](https://raw.githubusercontent.com/snyk/agent-scan/main/README.md), read 2026-11-16. - **snyk-19** (vendor): Snyk's MCP Governance docs say Evo can block MCP server execution on the developer's machine at invocation.Quote: "can block server execution on the developer's machine at invocation"Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md), read 2026-11-16. - **snyk-20** (vendor): Snyk's docs say that when an agent calls an unauthorized MCP server, Evo halts execution on the developer's workstation and notifies the developer of the policy, the blocked tool call and the server name.Quote: "Evo halts execution inline on the developer's workstation."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md), read 2026-11-16. - **snyk-21** (vendor): Snyk lists the agents covered by Agent Behavior Governance as Claude Code, Codex, Cursor and GitHub Copilot, a narrower list than the discovery list; Codex in user mode needs the developer to accept a prompt.Quote: "Claude Code, Codex, Cursor, and GitHub Copilot."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance.md), read 2026-11-16. - **snyk-22** (vendor): Snyk says Agent Behavior Governance evaluates each action against policy before it runs, using full-session context rather than single commands.Quote: "evaluating each action against the policy before it runs."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance.md), read 2026-11-16. - **snyk-23** (none-found): None found on 2026-11-16: the MCP Governance page describes allow and block decisions by server and risk, and does not describe inspecting MCP tool-call arguments or results.Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md), read 2026-11-16. - **snyk-24** (vendor): Snyk's docs describe an Allowed MCP Servers classification where unlisted servers show as Not Allowed; the allowlist ships empty and disabled.Quote: "Classify the servers you trust. Everything else shows as Not Allowed"Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md), read 2026-11-16. - **snyk-25** (vendor): Snyk's docs say policies can use risk dimensions such as prompt injection in a tool, destructive capabilities or unofficial publisher status, with a block preview before saving.Quote: "flag servers with prompt injection in a tool, destructive capabilities, or unofficial publisher status."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md), read 2026-11-16. - **snyk-26** (vendor): On 2026-09-30 Snyk announced Govern Agent Behavior is generally available, starting with MCP Governance; it says Evo ADS can log unauthorized MCP usage for audit or block it on the endpoint.Quote: "Evo ADS can log unauthorized MCP usage for audit, or block it outright, directly on the endpoint"Source: [snyk.io](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/), read 2026-11-16. - **snyk-27** (vendor): Snyk documents limits for Agent Behavior Governance: 5k hook calls per day per machine and a fixed seven-day data retention.Quote: "Data is retained for seven days. The data retention setting is fixed at seven days and cannot be changed."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance.md), read 2026-11-16. - **snyk-28** (none-found): None found on 2026-11-16: the MCP Governance docs page does not describe export of the MCP audit trail to a SIEM or scoping of policy to a team or group.Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md), read 2026-11-16. - **snyk-29** (vendor): Snyk's docs say ADS is installed on one machine or fleet-wide through an MDM tool, with Jamf Pro for macOS and Microsoft Intune for Windows.Quote: "Choose which products to roll out, then install on a single machine or across your company through your mobile device management (MDM) tool."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/activation-and-deployment.md), read 2026-11-16. - **snyk-30** (vendor): Snyk's docs list the supported OS minimums: Linux (Ubuntu 22.04, Debian 12, Amazon Linux 2023, Fedora 40), Windows 10 or later and macOS 15 or later.Quote: "macOS (Apple silicon and Intel): 15 or later"Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/activation-and-deployment/install-on-a-single-machine.md), read 2026-11-16. - **snyk-31** (vendor): Snyk's docs say the first install run can take several minutes; no fixed time-to-deploy figure is published.Quote: "The first run can take several minutes while the installer downloads the products your Tenant has selected."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/activation-and-deployment/install-on-a-single-machine.md), read 2026-11-16. - **snyk-32** (vendor): Snyk's docs say the scheduled discovery scan runs every 24 hours by default on macOS and Windows.Quote: "scans run every 24 hours by default on macOS and Windows."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/activation-and-deployment.md), read 2026-11-16. - **snyk-33** (vendor): The Agent Scan README shows a command-line scan via uvx and says a background MDM mode reports to Snyk Evo.Quote: "uvx snyk-agent-scan@latest"Source: [raw.githubusercontent.com](https://raw.githubusercontent.com/snyk/agent-scan/main/README.md), read 2026-11-16. - **snyk-34** (vendor): Snyk warns that Agent Scan CLI output is experimental and that it does not recommend production workflows built on specific output fields.Quote: "We do not recommend building production workflows that depend on specific CLI output fields or risk names."Source: [raw.githubusercontent.com](https://raw.githubusercontent.com/snyk/agent-scan/main/README.md), read 2026-11-16. - **snyk-35** (vendor): Snyk's pricing page publishes a rate of 1.0 credit per active machine per day for Evo ADS Coding Agent Security, with 1 credit equal to $1.Quote: "Evo ADS - Coding Agent Security ? per Active Machine / per Day 1.0 credits"Source: [snyk.io](https://snyk.io/plans/), read 2026-11-16. - **snyk-36** (vendor): Snyk says Evo capabilities such as Coding Agent security are not on the Free or Team plans and need an Enterprise Platform Subscription with sales contact.Quote: "requires an Enterprise Platform Subscription"Source: [snyk.io](https://snyk.io/plans/), read 2026-11-16. - **snyk-37** (vendor): Snyk's Evo docs are public without sign-in and describe ADS as Agent Supply Chain Security, Agent Behavior Governance and Trusted Output Assurance (Snyk Studio).Quote: "ADS secures three parts of that activity: what agents use, what they do, and what they generate."Source: [docs.snyk.io](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads.md), read 2026-11-16. - **snyk-38** (vendor): On 2026-06-23 Snyk published telemetry from nearly 9,700 developer environments, reporting that 1 in 12 developers with MCP servers has a high or critical finding.Quote: "One in 12 developers with MCP servers has a high or critical finding."Source: [snyk.io](https://snyk.io/news/snyk-launches-evo-agentic-development-security/), read 2026-11-16. - **snyk-39** (vendor): Snyk's ADS page quotes Brendan Putek, Director of DevOps at Relay Network, as a named customer reference.Quote: "Working with Snyk, we landed on what I think is the right agentic development security architecture"Source: [snyk.io](https://snyk.io/evo/agentic-development-security/), read 2026-11-16. - **snyk-40** (vendor): On 2026-11-04 Snyk said it was named a Leader in The Forrester Wave: Agentic Development Security Tools, Q4 2026 (Snyk's claim).Quote: "named a Leader in The Forrester Wave™: Agentic Development Security Tools, Q4 2026 report."Source: [snyk.io](https://snyk.io/news/snyk-named-leader-forrester-wave-agentic-development-security-tools/), read 2026-11-16. - **snyk-41** (third-party): SiliconANGLE reported on 2026-06-23 that Evo ADS vets the MCP servers an agent uses and enforces policy on what an agent does as it runs.Quote: "monitors and enforces policy on what an agent does as it runs"Source: [siliconangle.com](https://siliconangle.com/2026/06/23/snyk-launches-evo-agentic-development-security-police-ai-coding-agents/), read 2026-11-16. - **snyk-42** (third-party): Globes reported on 2021-09-09 that Snyk raised $300 million at an $8.5 billion valuation, co-led by Sands Capital Ventures and Tiger Global.Quote: "Snyk has announced the completion of a $300 million financing round co-led by Sands Capital Ventures and Tiger Global at a company valuation of $8.5 billion."Source: [en.globes.co.il](https://en.globes.co.il/en/article-cybersecurity-co-snyk-raises-300m-at-85b-valuation-1001384232), read 2026-11-16. - **snyk-43** (vendor): On 2026-06-23 Snyk announced Evo Agentic Development Security, saying it blocks destructive actions before they execute and vets MCP servers, skills and tools.Quote: "Monitors and enforces real-time policy on what agents do while they operate"Source: [snyk.io](https://snyk.io/news/snyk-launches-evo-agentic-development-security/), read 2026-11-16. ## News mentioning Snyk (Agent Scan, Evo ADS) - [Snyk makes Evo ADS MCP Governance generally available](/news/snyk-evo-ads-mcp-governance-generally-available/) ## Related pages - [All vendor profiles](/vendors/) - [SentinelOne (Prompt Security) profile](/vendors/sentinelone/) - [Zenity profile](/vendors/zenity/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Zenity MCP security: live tool call control URL: https://mcpservercybersecurity.com/vendors/zenity/ Updated: 2026-11-16 Zenity says its MCP Security product gives a live view of every MCP request and tool call and can allow, modify or block a call based on live context, through native hooks and its gateway. It imports servers from its inventory or public registries so known and shadow servers appear together, and it names Claude Code and Cowork, ChatGPT Codex, GitHub Copilot and Cursor as the coding agents where it enforces boundaries. Zenity publishes MCP research, and third-party press confirmed a $125 million funding round. Gaps on 16 Nov 2026: pre-run scanning of a server, a supported operating system list, the endpoint install method and time to deploy, readable documentation and a price are none found. ## Strength and weakness for MCP Server Security **Strength:** Offers a live view of every MCP request and tool call with the ability to allow, modify or block calls, and names the coding clients it covers (zenity-18, zenity-24, zenity-20). **Weakness:** Discovery and control route through hooks and a gateway; pre-run server scanning, an OS list and deploy time are none found on 2026-11-16 (zenity-22, zenity-29). ## Scores and evidence Zenity is 11th in the overall ranking at 44.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Zenity: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Discovery | 5 of 10 | 25 | The vendor claims discovery of coding agents and their MCP servers and tools on laptops and names Claude Code, Codex, Copilot and Cursor; vendor claims only. | [zenity-19](#fact-zenity-19), [zenity-20](#fact-zenity-20), [zenity-18](#fact-zenity-18), [zenity-21](#fact-zenity-21) | | Vetting before use | 3 of 10 | 20 | The vendor describes tool poisoning governance and sandboxing of skills; a pre-run server scan or registry: none found. | [zenity-22](#fact-zenity-22), [zenity-23](#fact-zenity-23), [zenity-17](#fact-zenity-17) | | Runtime control of tool calls | 6 of 10 | 20 | The vendor claims allow, modify or block of tool calls through native hooks and its gateway; argument inspection detail: none found. | [zenity-24](#fact-zenity-24), [zenity-50](#fact-zenity-50), [zenity-51](#fact-zenity-51) | | Deployment | 3 of 10 | 15 | Gateway URL registration and native hooks are named; OS, endpoint install method and time to deploy: none found. | [zenity-25](#fact-zenity-25), [zenity-26](#fact-zenity-26), [zenity-29](#fact-zenity-29), [zenity-58](#fact-zenity-58) | | Governance and audit | 5 of 10 | 5 | The vendor claims a live view of requests and identities and audit trails; SIEM export and per-team policy: none found. | [zenity-27](#fact-zenity-27), [zenity-28](#fact-zenity-28) | | Transparency | 3 of 10 | 10 | The vendor publishes MCP research; readable documentation and a price: none found. | [zenity-15](#fact-zenity-15), [zenity-16](#fact-zenity-16), [zenity-39](#fact-zenity-39), [zenity-59](#fact-zenity-59) | | Proof | 7 of 10 | 5 | Third-party press confirms a $125 million round and a named customer is listed; analyst placements are vendor claims. | [zenity-05](#fact-zenity-05), [zenity-02](#fact-zenity-02), [zenity-04](#fact-zenity-04), [zenity-08](#fact-zenity-08) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 11th | 44.0 | | [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) | 11th | 45.0 | | [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) | 13th | 42.5 | | [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) | joint 11th | 41.5 | | [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) | joint 10th | 49.0 | | [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) | 11th | 47.5 | ## Product, deployment, price and owner - Product assessed: Zenity MCP Security (live view and gateway-based control of MCP servers and tools) plus Zenity coding and personal agent security for Claude Code, Cursor, GitHub Copilot and ChatGPT Codex - Deployment: Connects to platforms by API and integration (Microsoft, Salesforce, OpenAI, Google, AWS, Okta and others per docs navigation); coding agents via native hooks, OpenTelemetry and the Zenity MCP gateway; no agent or proxy install details published - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 - Owner: Zenity (standalone, venture-backed company; no parent company stated on its pages) - Website: [zenity.io](https://zenity.io) ## Facts and sources - **zenity-01** (vendor): Zenity describes itself as an AI agent security platform that spans discovery, policy, identity, and runtime defense.Quote: "Zenity secures long-horizon AI agents at the decision, the exact point where context and intent come together to create the highest risk."Source: [zenity.io](https://zenity.io/), read 2026-11-16. - **zenity-02** (vendor): On 2026-08-03, Zenity announced a $125 million Series C led by Norwest; it is a standalone venture-backed company headquartered in New York.Quote: "today announced a $125 million Series C led by Norwest."Source: [zenity.io](https://zenity.io/press-release/zenity-raises-125-million-to-secure-the-era-of-1-billion-ai-agents), read 2026-11-16. - **zenity-03** (vendor): Zenity lists its funding history on its company page: a $5M seed in 2021, a $16.5M Series A in 2023 and a $38M Series B in 2024.Quote: "Zenity raises $38M Series B round of fundraising, co-led by Third Point Ventures and DTCP"Source: [zenity.io](https://zenity.io/company), read 2026-11-16. - **zenity-04** (vendor): Zenity says most of its customers are Fortune 500 and Global 2000 organizations and names SoftBank Corp. as a customer.Quote: "including SoftBank Corp."Source: [zenity.io](https://zenity.io/press-release/zenity-raises-125-million-to-secure-the-era-of-1-billion-ai-agents), read 2026-11-16. - **zenity-05** (third-party): Israeli business outlet Calcalist CTech, on 2026-09-03, reports that Zenity recently raised $125 million in a round led by Norwest.Quote: "a company that recently raised $125 million in a funding round led by Norwest"Source: [calcalistech.com](https://www.calcalistech.com/ctechnews/article/rsdllsuy3), read 2026-11-16. - **zenity-06** (vendor): On 2026-09-15, Zenity announced recognition in two independent analyst reports, from Latio and Software Analyst Cyber Research (SACR).Quote: "today announced recognition in two independent analyst reports examining the evolving AI agent security market."Source: [zenity.io](https://zenity.io/press-release/zenity-earns-recognition-in-ai-agent-security-reports-from-latio-and-sacr), read 2026-11-16. - **zenity-07** (vendor): Zenity says SACR places it among vendors combining runtime behavior analysis, MCP and tool governance, agent identity, audit evidence and response.Quote: "SACR identifies Zenity among vendors that combine multiple agent security capabilities, including runtime behavior analysis, MCP and tool governance"Source: [zenity.io](https://zenity.io/press-release/zenity-earns-recognition-in-ai-agent-security-reports-from-latio-and-sacr), read 2026-11-16. - **zenity-08** (vendor): On 2026-09-23, Zenity announced it was named a Market Shaper in Gartner's first Emerging Market Quadrant for AI Application Security (startup vendors).Quote: "Zenity has been named a Market Shaper in Gartner"Source: [zenity.io](https://zenity.io/company/newsroom/press-releases), read 2026-11-16. - **zenity-09** (vendor): Zenity lists a Gartner paper naming it the "Company to Beat" in AI agent governance on its platform pages.Quote: "Gartner® Names Zenity as the Company to Beat in AI Agent Governance"Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-15** (vendor): Zenity publishes a 2025 MCP Server Security Report on MCP vulnerabilities, threat patterns and governance gaps.Quote: "dives into the real-world vulnerabilities, threat patterns, and governance gaps behind this fast-growing standard"Source: [zenity.io](https://zenity.io/resources/white-papers/mcp-server-security-report), read 2026-11-16. - **zenity-16** (vendor): On 2026-09-29, Zenity published a blog post arguing that MCP spec changes need a separate runtime layer that judges intent.Quote: "Treat this spec wave as an infrastructure upgrade, and budget separately for a runtime layer that judges intent, not just authentication."Source: [zenity.io](https://zenity.io/blog/mcp-is-growing-up), read 2026-11-16. - **zenity-17** (vendor): On 2026-10-14, Zenity published a blog post treating the MCP server as a supply chain risk.Quote: "the Model Context Protocol server has become the single most consequential dependency in that category"Source: [zenity.io](https://zenity.io/blog/mcp-server-supply-chain-risk), read 2026-11-16. - **zenity-18** (vendor): Zenity says its MCP Security product imports servers from its inventory or public registries so known and shadow servers show in one place.Quote: "Import and discovery: Add servers from your Zenity inventory or from public registries, so both known and shadow servers surface in one place."Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-19** (vendor): Zenity says it discovers coding and personal agents on developer laptops and desktops, plus the MCP servers, tools and data they connect to.Quote: "Discover every coding and personal agent across developer laptops and desktops, and the MCP servers, tools, and data they connect to."Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-20** (vendor): Zenity names Claude Code and Cowork, ChatGPT Codex, GitHub Copilot and Cursor as the coding and personal agents where it enforces boundaries.Quote: "enforces hard boundaries across Anthropic's Claude Code and Cowork, ChatGPT Codex, GitHub Copilot, and Cursor."Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-21** (vendor): Zenity says AI Observability extends visibility to coding assistants and local MCPs running outside centrally managed platforms.Quote: "AI Observability extends visibility to coding assistants and local MCPs running outside centrally managed platforms."Source: [zenity.io](https://zenity.io/platform/ai-observability), read 2026-11-16. - **zenity-22** (vendor): Zenity describes tool poisoning and tool drift (rug pull) as MCP risks its product governs; it does not document a pre-run code scan of servers.Quote: "Tool poisoning A malicious or manipulated tool description steers an agent into leaking data or taking an action it was never asked to."Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-23** (vendor): Zenity says it runs risky skills, files and code in a sandbox to see what they actually do.Quote: "running risky skills, files, and code in a sandbox to see what they actually do, not just what they claim"Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-24** (vendor): Zenity says its MCP product can allow, modify or block a tool call based on live context.Quote: "Allow, modify, or block a tool call based on live context, not a static allow-list."Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-25** (vendor): Zenity says MCP traffic is controlled by publishing a server and registering one gateway URL, so agents keep working unchanged.Quote: "Publish a server, register one gateway URL, and agents keep working unchanged."Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-26** (vendor): Zenity says it collects coding-agent telemetry through native hooks and OpenTelemetry.Quote: "through native hooks and OpenTelemetry: each step an agent takes, the code and data it touches"Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-27** (vendor): Zenity says its MCP Live View shows each request, tool invoked and identity behind it in real time.Quote: "Live View shows each request, the tool invoked, and the identity behind it, in real time."Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-28** (vendor): Zenity says it provides unified visibility, audit trails and enforcement across coding and personal agents.Quote: "Prove governance across every coding and personal agent, with unified visibility, audit trails, and enforcement."Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-29** (none-found): None found on 2026-11-16: operating systems supported, endpoint install method or time to deploy on developer machines, searched on the coding and personal agents page, MCP security page and platform page.Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-37** (vendor): Zenity's public docs navigation lists integrations for OpenAI, Google, AWS, Salesforce, Anthropic, Cursor, GitHub, LiteLLM, Cribl Stream, Okta, CrowdStrike and Palo Alto Networks.Quote: "Gateways and custom agents LiteLLM Custom agents Cribl Stream ACS schema Identity and endpoint security Okta CrowdStrike Palo Alto Networks"Source: [docs.zenity.io](https://docs.zenity.io), read 2026-11-16. - **zenity-39** (none-found): None found on 2026-11-16: documentation body text readable without sign-in; docs.zenity.io/user-guide/* and /release-notes redirected to a login page, and only the navigation tree of the docs home page loaded.Source: [docs.zenity.io](https://docs.zenity.io/user-guide/inline-prevention), read 2026-11-16. - **zenity-42** (vendor): Zenity says inventory appears the moment a platform is connected, with no separate discovery step.Quote: "Full inventory the moment a platform connects, no separate discovery step required"Source: [zenity.io](https://zenity.io/use-cases/risk-type/shadow-ai), read 2026-11-16. - **zenity-50** (vendor): Zenity says that through native agent hooks and the Zenity MCP gateway it can block or modify a dangerous tool call before it executes.Quote: "it can block or modify a dangerous tool call before it executes."Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-51** (vendor): Zenity names secret exfiltration, malicious package installs and unsafe tool chaining as risks it detects and prevents on coding agents.Quote: "Detect and prevent risks like secret exfiltration, malicious package installs, unsafe tool chaining, and an agent working around a security control"Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-58** (none-found): None found on 2026-11-16: a published time to deploy for the platform, searched on the platform page, product pages and press releases.Source: [zenity.io](https://zenity.io/platform), read 2026-11-16. - **zenity-59** (none-found): None found on 2026-11-16: a published price or pricing page; https://zenity.io/pricing returned HTTP 404 and the sitemap lists no pricing page.Source: [zenity.io](https://zenity.io/pricing), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Snyk (Agent Scan, Evo ADS) profile](/vendors/snyk/) - [Backslash Security profile](/vendors/backslash-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Best MCP security tools by situation URL: https://mcpservercybersecurity.com/best-for/ Updated: 2026-11-16 Five situation pages re-rank the 14 tools with weights that fit a different job: hundreds of developers, an inventory this month, an approved server catalog, stopping tool calls, and a regulated audit. The evidence is the same in all five, only the weights change. ## 5 situations - [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/): Snyk (Agent Scan, Evo ADS) leads at 76.0; Bay is 10th. - [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/): Snyk (Agent Scan, Evo ADS) leads at 78.5; Bay is 10th. - [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/): Snyk (Agent Scan, Evo ADS) leads at 77.5; Bay is 9th. - [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/): Harmonic Security leads at 72.5; Bay is 8th. - [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/): Cisco leads at 73.0; Bay is 10th. ## Who should pick what With no particular situation, the published ranking applies and Snyk (Agent Scan, Evo ADS) comes first. For a specific situation the weights change, and so can the order: - **[An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/)**: pick Snyk (Agent Scan, Evo ADS) (76.0). Bay is 10th at 49.0. At scale the first problem is knowing which servers exist on which machines, and the tool has to deploy to every laptop. - **[A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/)**: pick Snyk (Agent Scan, Evo ADS) (78.5). Bay is 10th at 48.0. Speed to an inventory rewards discovery breadth and light deployment; control can follow. - **[A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/)**: pick Snyk (Agent Scan, Evo ADS) (77.5). Bay is 9th at 47.5. An approved list depends on assessing each server before use, so vetting weighs most. - **[A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/)**: pick Harmonic Security (72.5). Bay is 8th at 52.5. When a tool call can delete or send data, stopping it on the machine as it happens matters most. - **[A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/)**: pick Cisco (73.0). Bay is 10th at 52.5. Auditors want the record, so the audit log and per-team policy weigh most after discovery. ## Related pages - [The overall ranking](/ranking/) - [Weights per situation](/method/) - [Vendor profiles](/vendors/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Best MCP security tool for hundreds of developers URL: https://mcpservercybersecurity.com/best-for/many-developers/ Updated: 2026-11-16 For an engineering organization where hundreds of developers add MCP servers to IDEs, Snyk (Agent Scan, Evo ADS) ranks first in this view. Discovery weighs 30 and deployment 20, because at scale the first problem is knowing which servers sit on which machines and reaching every laptop. ## The short answer Snyk (Agent Scan, Evo ADS) leads the view "An engineering organization where hundreds of developers add MCP servers to their IDEs" at 76.0 of 100. Bay is 10th at 49.0. Snyk (Agent Scan, Evo ADS) is ahead mainly on discovery (8 against 4, worth 12.0 points at weight 30). ## Why these weights At scale the first problem is knowing which servers exist on which machines, and the tool has to deploy to every laptop. **Weights used in this view** | Discovery | Vetting before use | Runtime control of tool calls | Deployment | Governance and audit | Transparency | Proof | |---|---|---|---|---|---|---| | 30 | 15 | 20 | 20 | 5 | 5 | 5 | **An engineering organization where hundreds of developers add MCP servers to their IDEs: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Discovery (weight 30) | Vetting before use (weight 15) | Runtime control of tool calls (weight 20) | Deployment (weight 20) | Governance and audit (weight 5) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [Snyk (Agent Scan, Evo ADS)](/vendors/snyk/) | 76.0 | [8](/vendors/snyk/#evidence-discovery) | [9](/vendors/snyk/#evidence-vetting) | [6](/vendors/snyk/#evidence-runtime) | [8](/vendors/snyk/#evidence-deploy) | [4](/vendors/snyk/#evidence-governance) | [9](/vendors/snyk/#evidence-transparency) | [8](/vendors/snyk/#evidence-proof) | | 2nd | [Harmonic Security](/vendors/harmonic-security/) | 72.5 | [8](/vendors/harmonic-security/#evidence-discovery) | [5](/vendors/harmonic-security/#evidence-vetting) | [8](/vendors/harmonic-security/#evidence-runtime) | [8](/vendors/harmonic-security/#evidence-deploy) | [8](/vendors/harmonic-security/#evidence-governance) | [4](/vendors/harmonic-security/#evidence-transparency) | [6](/vendors/harmonic-security/#evidence-proof) | | 3rd | [Cisco](/vendors/cisco/) | 71.5 | [7](/vendors/cisco/#evidence-discovery) | [8](/vendors/cisco/#evidence-vetting) | [6](/vendors/cisco/#evidence-runtime) | [8](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-governance) | [7](/vendors/cisco/#evidence-transparency) | [6](/vendors/cisco/#evidence-proof) | | 4th | [Knostic](/vendors/knostic/) | 61.0 | [6](/vendors/knostic/#evidence-discovery) | [7](/vendors/knostic/#evidence-vetting) | [5](/vendors/knostic/#evidence-runtime) | [7](/vendors/knostic/#evidence-deploy) | [6](/vendors/knostic/#evidence-governance) | [6](/vendors/knostic/#evidence-transparency) | [5](/vendors/knostic/#evidence-proof) | | 5th | [Backslash Security](/vendors/backslash-security/) | 58.5 | [6](/vendors/backslash-security/#evidence-discovery) | [7](/vendors/backslash-security/#evidence-vetting) | [6](/vendors/backslash-security/#evidence-runtime) | [5](/vendors/backslash-security/#evidence-deploy) | [6](/vendors/backslash-security/#evidence-governance) | [4](/vendors/backslash-security/#evidence-transparency) | [6](/vendors/backslash-security/#evidence-proof) | | 6th | [Pillar Security](/vendors/pillar-security/) | 58.0 | [6](/vendors/pillar-security/#evidence-discovery) | [6](/vendors/pillar-security/#evidence-vetting) | [6](/vendors/pillar-security/#evidence-runtime) | [6](/vendors/pillar-security/#evidence-deploy) | [4](/vendors/pillar-security/#evidence-governance) | [4](/vendors/pillar-security/#evidence-transparency) | [6](/vendors/pillar-security/#evidence-proof) | | 7th | [Microsoft](/vendors/microsoft/) | 56.5 | [8](/vendors/microsoft/#evidence-discovery) | [2](/vendors/microsoft/#evidence-vetting) | [4](/vendors/microsoft/#evidence-runtime) | [6](/vendors/microsoft/#evidence-deploy) | [5](/vendors/microsoft/#evidence-governance) | [8](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 8th | [Lasso Security](/vendors/lasso-security/) | 56.0 | [6](/vendors/lasso-security/#evidence-discovery) | [6](/vendors/lasso-security/#evidence-vetting) | [6](/vendors/lasso-security/#evidence-runtime) | [4](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-governance) | [4](/vendors/lasso-security/#evidence-transparency) | [7](/vendors/lasso-security/#evidence-proof) | | 9th | [Noma Security](/vendors/noma-security/) | 52.0 | [6](/vendors/noma-security/#evidence-discovery) | [6](/vendors/noma-security/#evidence-vetting) | [5](/vendors/noma-security/#evidence-runtime) | [4](/vendors/noma-security/#evidence-deploy) | [5](/vendors/noma-security/#evidence-governance) | [3](/vendors/noma-security/#evidence-transparency) | [6](/vendors/noma-security/#evidence-proof) | | 10th | [Bay](/vendors/bay/) | 49.0 | [4](/vendors/bay/#evidence-discovery) | [4](/vendors/bay/#evidence-vetting) | [6](/vendors/bay/#evidence-runtime) | [6](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-governance) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 11th | [Zenity](/vendors/zenity/) | 45.0 | [5](/vendors/zenity/#evidence-discovery) | [3](/vendors/zenity/#evidence-vetting) | [6](/vendors/zenity/#evidence-runtime) | [3](/vendors/zenity/#evidence-deploy) | [5](/vendors/zenity/#evidence-governance) | [3](/vendors/zenity/#evidence-transparency) | [7](/vendors/zenity/#evidence-proof) | | 12th | [CrowdStrike](/vendors/crowdstrike/) | 42.5 | [4](/vendors/crowdstrike/#evidence-discovery) | [3](/vendors/crowdstrike/#evidence-vetting) | [3](/vendors/crowdstrike/#evidence-runtime) | [6](/vendors/crowdstrike/#evidence-deploy) | [5](/vendors/crowdstrike/#evidence-governance) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 13th | [Koi](/vendors/koi/) | 41.0 | [4](/vendors/koi/#evidence-discovery) | [5](/vendors/koi/#evidence-vetting) | [2](/vendors/koi/#evidence-runtime) | [6](/vendors/koi/#evidence-deploy) | [2](/vendors/koi/#evidence-governance) | [1](/vendors/koi/#evidence-transparency) | [8](/vendors/koi/#evidence-proof) | | 14th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 37.0 | [4](/vendors/sentinelone/#evidence-discovery) | [3](/vendors/sentinelone/#evidence-vetting) | [3](/vendors/sentinelone/#evidence-runtime) | [4](/vendors/sentinelone/#evidence-deploy) | [3](/vendors/sentinelone/#evidence-governance) | [4](/vendors/sentinelone/#evidence-transparency) | [6](/vendors/sentinelone/#evidence-proof) | ## What would change this ranking **Top pair.** Harmonic Security rising from 8 to 10 on discovery (weight 30, so each point moves a total by 3.0) would put Harmonic Security above Snyk (Agent Scan, Evo ADS); the gap is 3.5 points. **Bay's next place.** Bay rising from 4 to 6 on discovery (weight 30, so each point moves a total by 3.0) would put Bay above Noma Security; the gap is 3.0 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool is best for hundreds of developers adding MCP servers to their IDEs? Snyk (Agent Scan, Evo ADS) ranks first in this view. At scale the first problem is knowing which servers exist on which machines, so the view weighs Discovery at 30, Runtime control at 20, Deployment at 20, Vetting at 15, and Governance, Transparency and Proof at 5 each. The table above shows every total under these weights. A tool that reads many clients' config files and installs through MDM scores well here even if its blocking is narrow. ## What does a large developer population change? Each developer chooses a client and adds servers on their own, so no single list of clients is complete. A tool that names more clients per operating system leaves fewer blind spots, and deployment by MDM or an endpoint agent matters more than a polished console. Time to deploy is rarely documented. Where a vendor states minutes, that is a vendor statement, and the vendor page marks it as one. ## What would change this view? A published list of supported clients for Windows, macOS and Linux would raise several vendors' Discovery scores. So would a second source confirming a vendor's deployment claim, since vendor claims alone stop at 7. ## Related pages - [All situations](/best-for/) - [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) - [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Best MCP security tool for a fast MCP inventory URL: https://mcpservercybersecurity.com/best-for/fast-inventory/ Updated: 2026-11-16 For a security team that needs an MCP inventory across machines this month, Snyk (Agent Scan, Evo ADS) ranks first in this view. Discovery weighs 40 and deployment 30, so the order rewards broad detection and light installation while control and vetting count for little. ## The short answer Snyk (Agent Scan, Evo ADS) leads the view "A security team that needs an MCP inventory across machines this month" at 78.5 of 100. Bay is 10th at 48.0. Snyk (Agent Scan, Evo ADS) is ahead mainly on discovery (8 against 4, worth 16.0 points at weight 40). ## Why these weights Speed to an inventory rewards discovery breadth and light deployment; control can follow. **Weights used in this view** | Discovery | Vetting before use | Runtime control of tool calls | Deployment | Governance and audit | Transparency | Proof | |---|---|---|---|---|---|---| | 40 | 10 | 5 | 30 | 5 | 5 | 5 | **A security team that needs an MCP inventory across machines this month: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Discovery (weight 40) | Vetting before use (weight 10) | Runtime control of tool calls (weight 5) | Deployment (weight 30) | Governance and audit (weight 5) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [Snyk (Agent Scan, Evo ADS)](/vendors/snyk/) | 78.5 | [8](/vendors/snyk/#evidence-discovery) | [9](/vendors/snyk/#evidence-vetting) | [6](/vendors/snyk/#evidence-runtime) | [8](/vendors/snyk/#evidence-deploy) | [4](/vendors/snyk/#evidence-governance) | [9](/vendors/snyk/#evidence-transparency) | [8](/vendors/snyk/#evidence-proof) | | 2nd | [Harmonic Security](/vendors/harmonic-security/) | 74.0 | [8](/vendors/harmonic-security/#evidence-discovery) | [5](/vendors/harmonic-security/#evidence-vetting) | [8](/vendors/harmonic-security/#evidence-runtime) | [8](/vendors/harmonic-security/#evidence-deploy) | [8](/vendors/harmonic-security/#evidence-governance) | [4](/vendors/harmonic-security/#evidence-transparency) | [6](/vendors/harmonic-security/#evidence-proof) | | 3rd | [Cisco](/vendors/cisco/) | 73.5 | [7](/vendors/cisco/#evidence-discovery) | [8](/vendors/cisco/#evidence-vetting) | [6](/vendors/cisco/#evidence-runtime) | [8](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-governance) | [7](/vendors/cisco/#evidence-transparency) | [6](/vendors/cisco/#evidence-proof) | | 4th | [Microsoft](/vendors/microsoft/) | 63.5 | [8](/vendors/microsoft/#evidence-discovery) | [2](/vendors/microsoft/#evidence-vetting) | [4](/vendors/microsoft/#evidence-runtime) | [6](/vendors/microsoft/#evidence-deploy) | [5](/vendors/microsoft/#evidence-governance) | [8](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 5th | [Knostic](/vendors/knostic/) | 63.0 | [6](/vendors/knostic/#evidence-discovery) | [7](/vendors/knostic/#evidence-vetting) | [5](/vendors/knostic/#evidence-runtime) | [7](/vendors/knostic/#evidence-deploy) | [6](/vendors/knostic/#evidence-governance) | [6](/vendors/knostic/#evidence-transparency) | [5](/vendors/knostic/#evidence-proof) | | 6th | [Pillar Security](/vendors/pillar-security/) | 58.0 | [6](/vendors/pillar-security/#evidence-discovery) | [6](/vendors/pillar-security/#evidence-vetting) | [6](/vendors/pillar-security/#evidence-runtime) | [6](/vendors/pillar-security/#evidence-deploy) | [4](/vendors/pillar-security/#evidence-governance) | [4](/vendors/pillar-security/#evidence-transparency) | [6](/vendors/pillar-security/#evidence-proof) | | 7th | [Backslash Security](/vendors/backslash-security/) | 57.0 | [6](/vendors/backslash-security/#evidence-discovery) | [7](/vendors/backslash-security/#evidence-vetting) | [6](/vendors/backslash-security/#evidence-runtime) | [5](/vendors/backslash-security/#evidence-deploy) | [6](/vendors/backslash-security/#evidence-governance) | [4](/vendors/backslash-security/#evidence-transparency) | [6](/vendors/backslash-security/#evidence-proof) | | 8th | [Lasso Security](/vendors/lasso-security/) | 54.0 | [6](/vendors/lasso-security/#evidence-discovery) | [6](/vendors/lasso-security/#evidence-vetting) | [6](/vendors/lasso-security/#evidence-runtime) | [4](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-governance) | [4](/vendors/lasso-security/#evidence-transparency) | [7](/vendors/lasso-security/#evidence-proof) | | 9th | [Noma Security](/vendors/noma-security/) | 51.5 | [6](/vendors/noma-security/#evidence-discovery) | [6](/vendors/noma-security/#evidence-vetting) | [5](/vendors/noma-security/#evidence-runtime) | [4](/vendors/noma-security/#evidence-deploy) | [5](/vendors/noma-security/#evidence-governance) | [3](/vendors/noma-security/#evidence-transparency) | [6](/vendors/noma-security/#evidence-proof) | | 10th | [Bay](/vendors/bay/) | 48.0 | [4](/vendors/bay/#evidence-discovery) | [4](/vendors/bay/#evidence-vetting) | [6](/vendors/bay/#evidence-runtime) | [6](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-governance) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 11th | [CrowdStrike](/vendors/crowdstrike/) | 46.5 | [4](/vendors/crowdstrike/#evidence-discovery) | [3](/vendors/crowdstrike/#evidence-vetting) | [3](/vendors/crowdstrike/#evidence-runtime) | [6](/vendors/crowdstrike/#evidence-deploy) | [5](/vendors/crowdstrike/#evidence-governance) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 12th | [Koi](/vendors/koi/) | 45.5 | [4](/vendors/koi/#evidence-discovery) | [5](/vendors/koi/#evidence-vetting) | [2](/vendors/koi/#evidence-runtime) | [6](/vendors/koi/#evidence-deploy) | [2](/vendors/koi/#evidence-governance) | [1](/vendors/koi/#evidence-transparency) | [8](/vendors/koi/#evidence-proof) | | 13th | [Zenity](/vendors/zenity/) | 42.5 | [5](/vendors/zenity/#evidence-discovery) | [3](/vendors/zenity/#evidence-vetting) | [6](/vendors/zenity/#evidence-runtime) | [3](/vendors/zenity/#evidence-deploy) | [5](/vendors/zenity/#evidence-governance) | [3](/vendors/zenity/#evidence-transparency) | [7](/vendors/zenity/#evidence-proof) | | 14th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 39.0 | [4](/vendors/sentinelone/#evidence-discovery) | [3](/vendors/sentinelone/#evidence-vetting) | [3](/vendors/sentinelone/#evidence-runtime) | [4](/vendors/sentinelone/#evidence-deploy) | [3](/vendors/sentinelone/#evidence-governance) | [4](/vendors/sentinelone/#evidence-transparency) | [6](/vendors/sentinelone/#evidence-proof) | ## What would change this ranking **Top pair.** Harmonic Security rising from 8 to 10 on discovery (weight 40, so each point moves a total by 4.0) would put Harmonic Security above Snyk (Agent Scan, Evo ADS); the gap is 4.5 points. **Bay's next place.** Bay rising from 4 to 5 on discovery (weight 40, so each point moves a total by 4.0) would put Bay above Noma Security; the gap is 3.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool gives a fast MCP inventory? Snyk (Agent Scan, Evo ADS) ranks first in this view. Speed to an inventory rewards discovery breadth and light deployment, so the weights are Discovery 40, Deployment 30, Vetting 10, and Runtime, Governance, Transparency and Proof at 5 each. Control can follow once the list exists. Teams that need both should read the stop-tool-calls view next. ## What counts as an inventory? For scoring, an inventory names the servers, the tools each exposes and the client that launched each one, per machine. A count of servers with no client and no host is a mention, and scores 1 to 3. Open-source scanners can produce this from a single machine without a console. They give a quick first view, though a fleet needs the vendor's deployment path to repeat it. ## What would change this view? A documented fleet view with per-machine detail, or a stated list of clients read on each operating system, would lift the Discovery scores that are currently held at 4 to 6 by gaps the documentation shows. ## Related pages - [All situations](/best-for/) - [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) - [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Best MCP security tool for an approved server catalog URL: https://mcpservercybersecurity.com/best-for/approved-catalog/ Updated: 2026-11-16 For a platform team that wants an approved list of vetted MCP servers, Snyk (Agent Scan, Evo ADS) ranks first in this view. Vetting before use weighs 40, because an approved list is only as good as the assessment behind each entry, and an enforced list needs runtime control too. ## The short answer Snyk (Agent Scan, Evo ADS) leads the view "A platform team that wants an approved list of vetted MCP servers" at 77.5 of 100. Bay is 9th at 47.5. Snyk (Agent Scan, Evo ADS) is ahead mainly on vetting before use (9 against 4, worth 20.0 points at weight 40). ## Why these weights An approved list depends on assessing each server before use, so vetting weighs most. **Weights used in this view** | Discovery | Vetting before use | Runtime control of tool calls | Deployment | Governance and audit | Transparency | Proof | |---|---|---|---|---|---|---| | 15 | 40 | 15 | 10 | 10 | 5 | 5 | **A platform team that wants an approved list of vetted MCP servers: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Discovery (weight 15) | Vetting before use (weight 40) | Runtime control of tool calls (weight 15) | Deployment (weight 10) | Governance and audit (weight 10) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [Snyk (Agent Scan, Evo ADS)](/vendors/snyk/) | 77.5 | [8](/vendors/snyk/#evidence-discovery) | [9](/vendors/snyk/#evidence-vetting) | [6](/vendors/snyk/#evidence-runtime) | [8](/vendors/snyk/#evidence-deploy) | [4](/vendors/snyk/#evidence-governance) | [9](/vendors/snyk/#evidence-transparency) | [8](/vendors/snyk/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 74.0 | [7](/vendors/cisco/#evidence-discovery) | [8](/vendors/cisco/#evidence-vetting) | [6](/vendors/cisco/#evidence-runtime) | [8](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-governance) | [7](/vendors/cisco/#evidence-transparency) | [6](/vendors/cisco/#evidence-proof) | | 3rd | [Harmonic Security](/vendors/harmonic-security/) | 65.0 | [8](/vendors/harmonic-security/#evidence-discovery) | [5](/vendors/harmonic-security/#evidence-vetting) | [8](/vendors/harmonic-security/#evidence-runtime) | [8](/vendors/harmonic-security/#evidence-deploy) | [8](/vendors/harmonic-security/#evidence-governance) | [4](/vendors/harmonic-security/#evidence-transparency) | [6](/vendors/harmonic-security/#evidence-proof) | | 4th | [Knostic](/vendors/knostic/) | 63.0 | [6](/vendors/knostic/#evidence-discovery) | [7](/vendors/knostic/#evidence-vetting) | [5](/vendors/knostic/#evidence-runtime) | [7](/vendors/knostic/#evidence-deploy) | [6](/vendors/knostic/#evidence-governance) | [6](/vendors/knostic/#evidence-transparency) | [5](/vendors/knostic/#evidence-proof) | | 5th | [Backslash Security](/vendors/backslash-security/) | 62.0 | [6](/vendors/backslash-security/#evidence-discovery) | [7](/vendors/backslash-security/#evidence-vetting) | [6](/vendors/backslash-security/#evidence-runtime) | [5](/vendors/backslash-security/#evidence-deploy) | [6](/vendors/backslash-security/#evidence-governance) | [4](/vendors/backslash-security/#evidence-transparency) | [6](/vendors/backslash-security/#evidence-proof) | | 6th | [Lasso Security](/vendors/lasso-security/) | 58.5 | [6](/vendors/lasso-security/#evidence-discovery) | [6](/vendors/lasso-security/#evidence-vetting) | [6](/vendors/lasso-security/#evidence-runtime) | [4](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-governance) | [4](/vendors/lasso-security/#evidence-transparency) | [7](/vendors/lasso-security/#evidence-proof) | | 7th | [Pillar Security](/vendors/pillar-security/) | 57.0 | [6](/vendors/pillar-security/#evidence-discovery) | [6](/vendors/pillar-security/#evidence-vetting) | [6](/vendors/pillar-security/#evidence-runtime) | [6](/vendors/pillar-security/#evidence-deploy) | [4](/vendors/pillar-security/#evidence-governance) | [4](/vendors/pillar-security/#evidence-transparency) | [6](/vendors/pillar-security/#evidence-proof) | | 8th | [Noma Security](/vendors/noma-security/) | 54.0 | [6](/vendors/noma-security/#evidence-discovery) | [6](/vendors/noma-security/#evidence-vetting) | [5](/vendors/noma-security/#evidence-runtime) | [4](/vendors/noma-security/#evidence-deploy) | [5](/vendors/noma-security/#evidence-governance) | [3](/vendors/noma-security/#evidence-transparency) | [6](/vendors/noma-security/#evidence-proof) | | 9th | [Bay](/vendors/bay/) | 47.5 | [4](/vendors/bay/#evidence-discovery) | [4](/vendors/bay/#evidence-vetting) | [6](/vendors/bay/#evidence-runtime) | [6](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-governance) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 10th | [Microsoft](/vendors/microsoft/) | 44.0 | [8](/vendors/microsoft/#evidence-discovery) | [2](/vendors/microsoft/#evidence-vetting) | [4](/vendors/microsoft/#evidence-runtime) | [6](/vendors/microsoft/#evidence-deploy) | [5](/vendors/microsoft/#evidence-governance) | [8](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | joint 11th | [Koi](/vendors/koi/) | 41.5 | [4](/vendors/koi/#evidence-discovery) | [5](/vendors/koi/#evidence-vetting) | [2](/vendors/koi/#evidence-runtime) | [6](/vendors/koi/#evidence-deploy) | [2](/vendors/koi/#evidence-governance) | [1](/vendors/koi/#evidence-transparency) | [8](/vendors/koi/#evidence-proof) | | joint 11th | [Zenity](/vendors/zenity/) | 41.5 | [5](/vendors/zenity/#evidence-discovery) | [3](/vendors/zenity/#evidence-vetting) | [6](/vendors/zenity/#evidence-runtime) | [3](/vendors/zenity/#evidence-deploy) | [5](/vendors/zenity/#evidence-governance) | [3](/vendors/zenity/#evidence-transparency) | [7](/vendors/zenity/#evidence-proof) | | 13th | [CrowdStrike](/vendors/crowdstrike/) | 39.0 | [4](/vendors/crowdstrike/#evidence-discovery) | [3](/vendors/crowdstrike/#evidence-vetting) | [3](/vendors/crowdstrike/#evidence-runtime) | [6](/vendors/crowdstrike/#evidence-deploy) | [5](/vendors/crowdstrike/#evidence-governance) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 14th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 34.5 | [4](/vendors/sentinelone/#evidence-discovery) | [3](/vendors/sentinelone/#evidence-vetting) | [3](/vendors/sentinelone/#evidence-runtime) | [4](/vendors/sentinelone/#evidence-deploy) | [3](/vendors/sentinelone/#evidence-governance) | [4](/vendors/sentinelone/#evidence-transparency) | [6](/vendors/sentinelone/#evidence-proof) | ## What would change this ranking **Top pair.** Cisco rising from 8 to 9 on vetting before use (weight 40, so each point moves a total by 4.0) would put Cisco above Snyk (Agent Scan, Evo ADS); the gap is 3.5 points. **Bay's next place.** Bay rising from 4 to 6 on vetting before use (weight 40, so each point moves a total by 4.0) would put Bay above Noma Security; the gap is 6.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool suits an approved catalog of MCP servers? Snyk (Agent Scan, Evo ADS) ranks first in this view. An approved list depends on assessing each server before use, so Vetting weighs 40, with Discovery 15, Runtime 15, Deployment 10, Governance 10, and Transparency and Proof 5 each. Vetting scores reward scanning code, tool descriptions and permissions, plus provenance signals and an enforced list. A scan with no way to enforce the verdict scores lower than one that blocks an unlisted server. ## What does a catalog need beyond a scan? A catalog needs a named approver, a record of what was reviewed and a trigger for review when a server changes. The MCP Registry verifies who owns a namespace but leaves security scanning of server code to others, so the review stays with you. The building-an-approved-catalog guide covers the approval workflow without reference to any product. ## What would change this view? Documented handling of servers that are not in a vendor's catalog, such as locally written stdio servers, would raise Vetting scores. So would an enforced allowlist that ships enabled rather than empty. ## Related pages - [All situations](/best-for/) - [A security team that needs an MCP inventory across machines this month](/best-for/fast-inventory/) - [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Best MCP security tool to stop tool calls URL: https://mcpservercybersecurity.com/best-for/stop-tool-calls/ Updated: 2026-11-16 For a team whose developers run agents with shell and file access through MCP tools, Harmonic Security ranks first in this view. Runtime control of tool calls weighs 40, because when a call can delete or send data, stopping it on the machine as it happens matters most. ## The short answer Harmonic Security leads the view "A team whose developers run agents with shell and file access through MCP tools" at 72.5 of 100. Bay is 8th at 52.5. Harmonic Security is ahead mainly on runtime control of tool calls (8 against 6, worth 8.0 points at weight 40). ## Why these weights When a tool call can delete or send data, stopping it on the machine as it happens matters most. **Weights used in this view** | Discovery | Vetting before use | Runtime control of tool calls | Deployment | Governance and audit | Transparency | Proof | |---|---|---|---|---|---|---| | 15 | 15 | 40 | 10 | 10 | 5 | 5 | **A team whose developers run agents with shell and file access through MCP tools: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Discovery (weight 15) | Vetting before use (weight 15) | Runtime control of tool calls (weight 40) | Deployment (weight 10) | Governance and audit (weight 10) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [Harmonic Security](/vendors/harmonic-security/) | 72.5 | [8](/vendors/harmonic-security/#evidence-discovery) | [5](/vendors/harmonic-security/#evidence-vetting) | [8](/vendors/harmonic-security/#evidence-runtime) | [8](/vendors/harmonic-security/#evidence-deploy) | [8](/vendors/harmonic-security/#evidence-governance) | [4](/vendors/harmonic-security/#evidence-transparency) | [6](/vendors/harmonic-security/#evidence-proof) | | 2nd | [Snyk (Agent Scan, Evo ADS)](/vendors/snyk/) | 70.0 | [8](/vendors/snyk/#evidence-discovery) | [9](/vendors/snyk/#evidence-vetting) | [6](/vendors/snyk/#evidence-runtime) | [8](/vendors/snyk/#evidence-deploy) | [4](/vendors/snyk/#evidence-governance) | [9](/vendors/snyk/#evidence-transparency) | [8](/vendors/snyk/#evidence-proof) | | 3rd | [Cisco](/vendors/cisco/) | 69.0 | [7](/vendors/cisco/#evidence-discovery) | [8](/vendors/cisco/#evidence-vetting) | [6](/vendors/cisco/#evidence-runtime) | [8](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-governance) | [7](/vendors/cisco/#evidence-transparency) | [6](/vendors/cisco/#evidence-proof) | | 4th | [Backslash Security](/vendors/backslash-security/) | 59.5 | [6](/vendors/backslash-security/#evidence-discovery) | [7](/vendors/backslash-security/#evidence-vetting) | [6](/vendors/backslash-security/#evidence-runtime) | [5](/vendors/backslash-security/#evidence-deploy) | [6](/vendors/backslash-security/#evidence-governance) | [4](/vendors/backslash-security/#evidence-transparency) | [6](/vendors/backslash-security/#evidence-proof) | | 5th | [Lasso Security](/vendors/lasso-security/) | 58.5 | [6](/vendors/lasso-security/#evidence-discovery) | [6](/vendors/lasso-security/#evidence-vetting) | [6](/vendors/lasso-security/#evidence-runtime) | [4](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-governance) | [4](/vendors/lasso-security/#evidence-transparency) | [7](/vendors/lasso-security/#evidence-proof) | | 6th | [Knostic](/vendors/knostic/) | 58.0 | [6](/vendors/knostic/#evidence-discovery) | [7](/vendors/knostic/#evidence-vetting) | [5](/vendors/knostic/#evidence-runtime) | [7](/vendors/knostic/#evidence-deploy) | [6](/vendors/knostic/#evidence-governance) | [6](/vendors/knostic/#evidence-transparency) | [5](/vendors/knostic/#evidence-proof) | | 7th | [Pillar Security](/vendors/pillar-security/) | 57.0 | [6](/vendors/pillar-security/#evidence-discovery) | [6](/vendors/pillar-security/#evidence-vetting) | [6](/vendors/pillar-security/#evidence-runtime) | [6](/vendors/pillar-security/#evidence-deploy) | [4](/vendors/pillar-security/#evidence-governance) | [4](/vendors/pillar-security/#evidence-transparency) | [6](/vendors/pillar-security/#evidence-proof) | | 8th | [Bay](/vendors/bay/) | 52.5 | [4](/vendors/bay/#evidence-discovery) | [4](/vendors/bay/#evidence-vetting) | [6](/vendors/bay/#evidence-runtime) | [6](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-governance) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 9th | [Noma Security](/vendors/noma-security/) | 51.5 | [6](/vendors/noma-security/#evidence-discovery) | [6](/vendors/noma-security/#evidence-vetting) | [5](/vendors/noma-security/#evidence-runtime) | [4](/vendors/noma-security/#evidence-deploy) | [5](/vendors/noma-security/#evidence-governance) | [3](/vendors/noma-security/#evidence-transparency) | [6](/vendors/noma-security/#evidence-proof) | | joint 10th | [Microsoft](/vendors/microsoft/) | 49.0 | [8](/vendors/microsoft/#evidence-discovery) | [2](/vendors/microsoft/#evidence-vetting) | [4](/vendors/microsoft/#evidence-runtime) | [6](/vendors/microsoft/#evidence-deploy) | [5](/vendors/microsoft/#evidence-governance) | [8](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | joint 10th | [Zenity](/vendors/zenity/) | 49.0 | [5](/vendors/zenity/#evidence-discovery) | [3](/vendors/zenity/#evidence-vetting) | [6](/vendors/zenity/#evidence-runtime) | [3](/vendors/zenity/#evidence-deploy) | [5](/vendors/zenity/#evidence-governance) | [3](/vendors/zenity/#evidence-transparency) | [7](/vendors/zenity/#evidence-proof) | | 12th | [CrowdStrike](/vendors/crowdstrike/) | 39.0 | [4](/vendors/crowdstrike/#evidence-discovery) | [3](/vendors/crowdstrike/#evidence-vetting) | [3](/vendors/crowdstrike/#evidence-runtime) | [6](/vendors/crowdstrike/#evidence-deploy) | [5](/vendors/crowdstrike/#evidence-governance) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 13th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 34.5 | [4](/vendors/sentinelone/#evidence-discovery) | [3](/vendors/sentinelone/#evidence-vetting) | [3](/vendors/sentinelone/#evidence-runtime) | [4](/vendors/sentinelone/#evidence-deploy) | [3](/vendors/sentinelone/#evidence-governance) | [4](/vendors/sentinelone/#evidence-transparency) | [6](/vendors/sentinelone/#evidence-proof) | | 14th | [Koi](/vendors/koi/) | 34.0 | [4](/vendors/koi/#evidence-discovery) | [5](/vendors/koi/#evidence-vetting) | [2](/vendors/koi/#evidence-runtime) | [6](/vendors/koi/#evidence-deploy) | [2](/vendors/koi/#evidence-governance) | [1](/vendors/koi/#evidence-transparency) | [8](/vendors/koi/#evidence-proof) | ## What would change this ranking **Top pair.** Snyk (Agent Scan, Evo ADS) rising from 6 to 7 on runtime control of tool calls (weight 40, so each point moves a total by 4.0) would put Snyk (Agent Scan, Evo ADS) above Harmonic Security; the gap is 2.5 points. **Bay's next place.** Bay rising from 6 to 8 on runtime control of tool calls (weight 40, so each point moves a total by 4.0) would put Bay above Pillar Security; the gap is 4.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool stops MCP tool calls? Harmonic Security ranks first in this view. When a tool call can delete or send data, stopping it on the machine as it happens matters most, so Runtime weighs 40, with Discovery 15, Vetting 15, Deployment 10, Governance 10, and Transparency and Proof 5 each. Runtime scores reward allow, block or approval decisions on each MCP tool call with arguments inspected. Control that covers only a few clients, or only previews, is held to the lower anchors. ## Why do arguments matter? A rule that allows a tool by name cannot tell a harmless call from a harmful one made through the same tool. Inspecting arguments lets a policy block a delete outside the project folder while allowing reads inside it. For most vendors, argument inspection for MCP calls is none found on 16 Nov 2026, which keeps their Runtime scores in the 3 to 6 range. ## What would change this view? Generally available enforcement documented for the main clients, with the inspected fields listed, would raise Runtime scores. A second source confirming a vendor's blocking claim would lift it past the vendor-claim cap of 7. ## Related pages - [All situations](/best-for/) - [A platform team that wants an approved list of vetted MCP servers](/best-for/approved-catalog/) - [A regulated company that must show which MCP servers ran and what they did](/best-for/regulated-audit/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Best MCP security tool for a regulated audit URL: https://mcpservercybersecurity.com/best-for/regulated-audit/ Updated: 2026-11-16 For a regulated company that must show which MCP servers ran and what they did, Cisco ranks first in this view. Governance and audit weighs 30 and Discovery 25, because auditors want the record, the per-team policy and a complete list of servers. ## The short answer Cisco leads the view "A regulated company that must show which MCP servers ran and what they did" at 73.0 of 100. Bay is 10th at 52.5. Cisco is ahead mainly on discovery (7 against 4, worth 7.5 points at weight 25). ## Why these weights Auditors want the record, so the audit log and per-team policy weigh most after discovery. **Weights used in this view** | Discovery | Vetting before use | Runtime control of tool calls | Deployment | Governance and audit | Transparency | Proof | |---|---|---|---|---|---|---| | 25 | 15 | 15 | 5 | 30 | 5 | 5 | **A regulated company that must show which MCP servers ran and what they did: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Discovery (weight 25) | Vetting before use (weight 15) | Runtime control of tool calls (weight 15) | Deployment (weight 5) | Governance and audit (weight 30) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [Cisco](/vendors/cisco/) | 73.0 | [7](/vendors/cisco/#evidence-discovery) | [8](/vendors/cisco/#evidence-vetting) | [6](/vendors/cisco/#evidence-runtime) | [8](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-governance) | [7](/vendors/cisco/#evidence-transparency) | [6](/vendors/cisco/#evidence-proof) | | 2nd | [Harmonic Security](/vendors/harmonic-security/) | 72.5 | [8](/vendors/harmonic-security/#evidence-discovery) | [5](/vendors/harmonic-security/#evidence-vetting) | [8](/vendors/harmonic-security/#evidence-runtime) | [8](/vendors/harmonic-security/#evidence-deploy) | [8](/vendors/harmonic-security/#evidence-governance) | [4](/vendors/harmonic-security/#evidence-transparency) | [6](/vendors/harmonic-security/#evidence-proof) | | 3rd | [Snyk (Agent Scan, Evo ADS)](/vendors/snyk/) | 67.0 | [8](/vendors/snyk/#evidence-discovery) | [9](/vendors/snyk/#evidence-vetting) | [6](/vendors/snyk/#evidence-runtime) | [8](/vendors/snyk/#evidence-deploy) | [4](/vendors/snyk/#evidence-governance) | [9](/vendors/snyk/#evidence-transparency) | [8](/vendors/snyk/#evidence-proof) | | 4th | [Lasso Security](/vendors/lasso-security/) | 61.5 | [6](/vendors/lasso-security/#evidence-discovery) | [6](/vendors/lasso-security/#evidence-vetting) | [6](/vendors/lasso-security/#evidence-runtime) | [4](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-governance) | [4](/vendors/lasso-security/#evidence-transparency) | [7](/vendors/lasso-security/#evidence-proof) | | joint 5th | [Backslash Security](/vendors/backslash-security/) | 60.0 | [6](/vendors/backslash-security/#evidence-discovery) | [7](/vendors/backslash-security/#evidence-vetting) | [6](/vendors/backslash-security/#evidence-runtime) | [5](/vendors/backslash-security/#evidence-deploy) | [6](/vendors/backslash-security/#evidence-governance) | [4](/vendors/backslash-security/#evidence-transparency) | [6](/vendors/backslash-security/#evidence-proof) | | joint 5th | [Knostic](/vendors/knostic/) | 60.0 | [6](/vendors/knostic/#evidence-discovery) | [7](/vendors/knostic/#evidence-vetting) | [5](/vendors/knostic/#evidence-runtime) | [7](/vendors/knostic/#evidence-deploy) | [6](/vendors/knostic/#evidence-governance) | [6](/vendors/knostic/#evidence-transparency) | [5](/vendors/knostic/#evidence-proof) | | 7th | [Microsoft](/vendors/microsoft/) | 54.0 | [8](/vendors/microsoft/#evidence-discovery) | [2](/vendors/microsoft/#evidence-vetting) | [4](/vendors/microsoft/#evidence-runtime) | [6](/vendors/microsoft/#evidence-deploy) | [5](/vendors/microsoft/#evidence-governance) | [8](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | joint 8th | [Noma Security](/vendors/noma-security/) | 53.0 | [6](/vendors/noma-security/#evidence-discovery) | [6](/vendors/noma-security/#evidence-vetting) | [5](/vendors/noma-security/#evidence-runtime) | [4](/vendors/noma-security/#evidence-deploy) | [5](/vendors/noma-security/#evidence-governance) | [3](/vendors/noma-security/#evidence-transparency) | [6](/vendors/noma-security/#evidence-proof) | | joint 8th | [Pillar Security](/vendors/pillar-security/) | 53.0 | [6](/vendors/pillar-security/#evidence-discovery) | [6](/vendors/pillar-security/#evidence-vetting) | [6](/vendors/pillar-security/#evidence-runtime) | [6](/vendors/pillar-security/#evidence-deploy) | [4](/vendors/pillar-security/#evidence-governance) | [4](/vendors/pillar-security/#evidence-transparency) | [6](/vendors/pillar-security/#evidence-proof) | | 10th | [Bay](/vendors/bay/) | 52.5 | [4](/vendors/bay/#evidence-discovery) | [4](/vendors/bay/#evidence-vetting) | [6](/vendors/bay/#evidence-runtime) | [6](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-governance) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 11th | [Zenity](/vendors/zenity/) | 47.5 | [5](/vendors/zenity/#evidence-discovery) | [3](/vendors/zenity/#evidence-vetting) | [6](/vendors/zenity/#evidence-runtime) | [3](/vendors/zenity/#evidence-deploy) | [5](/vendors/zenity/#evidence-governance) | [3](/vendors/zenity/#evidence-transparency) | [7](/vendors/zenity/#evidence-proof) | | 12th | [CrowdStrike](/vendors/crowdstrike/) | 42.5 | [4](/vendors/crowdstrike/#evidence-discovery) | [3](/vendors/crowdstrike/#evidence-vetting) | [3](/vendors/crowdstrike/#evidence-runtime) | [6](/vendors/crowdstrike/#evidence-deploy) | [5](/vendors/crowdstrike/#evidence-governance) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 13th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 35.0 | [4](/vendors/sentinelone/#evidence-discovery) | [3](/vendors/sentinelone/#evidence-vetting) | [3](/vendors/sentinelone/#evidence-runtime) | [4](/vendors/sentinelone/#evidence-deploy) | [3](/vendors/sentinelone/#evidence-governance) | [4](/vendors/sentinelone/#evidence-transparency) | [6](/vendors/sentinelone/#evidence-proof) | | 14th | [Koi](/vendors/koi/) | 34.0 | [4](/vendors/koi/#evidence-discovery) | [5](/vendors/koi/#evidence-vetting) | [2](/vendors/koi/#evidence-runtime) | [6](/vendors/koi/#evidence-deploy) | [2](/vendors/koi/#evidence-governance) | [1](/vendors/koi/#evidence-transparency) | [8](/vendors/koi/#evidence-proof) | ## What would change this ranking **Top pair.** Harmonic Security rising from 8 to 9 on governance and audit (weight 30, so each point moves a total by 3.0) would put Harmonic Security above Cisco; the gap is 0.5 points. **Bay's next place.** Bay rising from 7 to 8 on governance and audit (weight 30, so each point moves a total by 3.0) would put Bay above Pillar Security; the gap is 0.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool suits a regulated audit of MCP servers? Cisco ranks first in this view. Auditors want the record, so Governance and audit weighs 30, with Discovery 25, Vetting 15, Runtime 15, Deployment 5, and Transparency and Proof 5 each. Governance scores reward policy per team or server, an audit log of servers and tool calls, and export to a SIEM. A log that lives only in the vendor console scores below one that exports in an open format. ## What does an auditor ask for? Typical requests are a list of servers in use on a date, who approved each, the calls made and who made them. The audit logging guide lists the fields to record and the trade-offs on retention. Retention differs by product. One vendor documents a fixed seven-day period for its agent governance data, so a longer record needs an export. ## What would change this view? Documented SIEM export and per-team scoping would raise several Governance scores that sit at 2 to 5 today. A published retention setting would help buyers who must keep records for a fixed period. ## Related pages - [All situations](/best-for/) - [A team whose developers run agents with shell and file access through MCP tools](/best-for/stop-tool-calls/) - [An engineering organization where hundreds of developers add MCP servers to their IDEs](/best-for/many-developers/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # MCP Risk Register URL: https://mcpservercybersecurity.com/mcp-risk-register/ Updated: 2026-11-16 The MCP Risk Register lists 42 risks for MCP servers in six dimensions: supply chain, authorization, transport and network, tool descriptions and schemas, execution and filesystem, and data exposure and logging. Each row has impact, controls, a written severity rule (19 high, 21 medium, 2 low) and primary sources read on 16 Nov 2026. ## All 42 items **MCP Risk Register: 42 items** | ID | Item | Category | Severity | |---|---|---|---| | [R-01](#R-01) | Look-alike package name (typosquatting) | [Supply chain](/mcp-risk-register/supply-chain/) | high | | [R-02](#R-02) | Repackaged server with altered endpoints | [Supply chain](/mcp-risk-register/supply-chain/) | high | | [R-03](#R-03) | Unreviewed code run through a package launcher | [Supply chain](/mcp-risk-register/supply-chain/) | high | | [R-04](#R-04) | Vulnerable or compromised dependency inside a server | [Supply chain](/mcp-risk-register/supply-chain/) | medium | | [R-05](#R-05) | Tool definitions change after approval (rug pull) | [Supply chain](/mcp-risk-register/supply-chain/) | medium | | [R-06](#R-06) | Registry listing mistaken for a security review | [Supply chain](/mcp-risk-register/supply-chain/) | medium | | [R-07](#R-07) | Unknown servers configured on developer machines | [Supply chain](/mcp-risk-register/supply-chain/) | high | | [R-08](#R-08) | Token passthrough and missing audience validation | [Authorization](/mcp-risk-register/authorization/) | high | | [R-09](#R-09) | HTTP endpoint reachable without authentication | [Authorization](/mcp-risk-register/authorization/) | high | | [R-10](#R-10) | Confused deputy at an MCP proxy server | [Authorization](/mcp-risk-register/authorization/) | medium | | [R-11](#R-11) | Client credentials sent to a server-chosen authorization server | [Authorization](/mcp-risk-register/authorization/) | high | | [R-12](#R-12) | Over-broad token scopes | [Authorization](/mcp-risk-register/authorization/) | medium | | [R-13](#R-13) | Authorization policy skipped on direct calls | [Authorization](/mcp-risk-register/authorization/) | medium | | [R-14](#R-14) | Authorization server mix-up and unverified issuer | [Authorization](/mcp-risk-register/authorization/) | medium | | [R-15](#R-15) | State handle or message delivered to the wrong client | [Authorization](/mcp-risk-register/authorization/) | medium | | [R-16](#R-16) | DNS rebinding against a local HTTP server | [Transport and network](/mcp-risk-register/transport-and-network/) | high | | [R-17](#R-17) | Local server listening on all interfaces | [Transport and network](/mcp-risk-register/transport-and-network/) | medium | | [R-18](#R-18) | Server-side request forgery through discovery and fetch tools | [Transport and network](/mcp-risk-register/transport-and-network/) | high | | [R-19](#R-19) | Credentials carried across cross-origin redirects | [Transport and network](/mcp-risk-register/transport-and-network/) | medium | | [R-20](#R-20) | Remote server without TLS or with unverified identity | [Transport and network](/mcp-risk-register/transport-and-network/) | medium | | [R-21](#R-21) | Unbounded requests and sessions exhaust a server | [Transport and network](/mcp-risk-register/transport-and-network/) | low | | [R-22](#R-22) | Unrestricted outbound access from server processes | [Transport and network](/mcp-risk-register/transport-and-network/) | medium | | [R-23](#R-23) | Poisoned tool description | [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) | high | | [R-24](#R-24) | Tool shadowing across servers | [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) | medium | | [R-25](#R-25) | Hostile content in schemas and return values | [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) | medium | | [R-26](#R-26) | Annotations trusted as enforcement | [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) | medium | | [R-27](#R-27) | Prompt injection through tool results | [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) | high | | [R-28](#R-28) | Server-supplied schemas that exhaust or confuse clients | [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) | low | | [R-29](#R-29) | Local server runs with the user's full privileges | [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) | high | | [R-30](#R-30) | Malicious startup command in configuration | [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) | high | | [R-31](#R-31) | Command injection through arguments or URL opening | [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) | high | | [R-32](#R-32) | Path traversal and arbitrary file read in file tools | [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) | high | | [R-33](#R-33) | Project configuration from a cloned repository | [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) | high | | [R-34](#R-34) | Tool calls auto-approved without full arguments | [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) | high | | [R-35](#R-35) | Developer tools that spawn processes on a port | [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) | medium | | [R-36](#R-36) | Secrets inherited through the environment | [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) | high | | [R-37](#R-37) | Tokens stored in plaintext | [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) | medium | | [R-38](#R-38) | Data exfiltration through legitimate tool calls | [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) | high | | [R-39](#R-39) | Sensitive data requested through form elicitation | [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) | medium | | [R-40](#R-40) | No record of tool calls | [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) | medium | | [R-41](#R-41) | Secrets and personal data written into logs | [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) | medium | | [R-42](#R-42) | Protected data returned through resources and templates | [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) | medium | ## What the register covers The register is a working list for engineers who must decide which MCP servers developers may install and how to constrain them. It follows the MCP specification's own security pages and the OWASP MCP Security Cheat Sheet, and adds advisories and release notes dated from August to November 2026. Rows describe the risk, the impact and the controls. They omit exploit code, payloads and bypass steps. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security). ## How the rows are organized There are 42 rows with identifiers R-01 to R-42. Each belongs to one of six dimensions and has a severity, controls an engineer can apply and at least one primary source. The companion MCP Security Checklist maps its steps to these identifiers. - Supply chain: 7 rows - Authorization: 8 rows - Transport and network: 7 rows - Tool descriptions and schemas: 6 rows - Execution and filesystem: 7 rows - Data exposure and logging: 7 rows Sources: [OWASP MCP Top 10](https://owasp.org/www-project-mcp-top-10/); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization). ## What changed in the protocol in 2026 The 2026-07-28 revision removes protocol-level sessions and deprecates Sampling, Roots, Logging and Dynamic Client Registration, with earliest removal in the first revision released on or after 2027-07-28. Rows that mention sessions or registration note the revision they apply to. SDK advisories published in October 2026 show where the specification's rules, such as token audience checks, needed code changes and, in some cases, a setting that upgrading alone does not enable. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23). ## Item detail ### R-01: Look-alike package name (typosquatting) A package is published under a name one keystroke away from a trusted MCP server, in the hope that a developer or a script installs it by mistake. **Impact:** The wrong package runs with the same permissions as the right one, including access to the user's files and environment variables. **Controls:** - Compare the exact package name, publisher and repository against the vendor's documentation before the first install. - Keep an approved server list and block installs of anything not on it (VS Code and Copilot enterprise-managed settings use allowedMcpServers and deniedMcpServers; Claude Code uses managed-mcp.json). - Pin the version once reviewed so a later publish under the same name is not picked up. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### R-02: Repackaged server with altered endpoints A package copies a legitimate server's open-source code but replaces authentication or API addresses, and the README does not mention the change. Knostic described one such npm package on 26 Oct 2026 and assessed it as suspected malicious credential redirection. **Impact:** Application credentials the user supplies as environment variables go to a destination the user did not choose. **Controls:** - Diff the package contents against the upstream repository for endpoint and authentication changes before approval. - Restrict outbound network access from MCP server processes to the hosts the server needs. - Rotate any credential that was supplied to a server later found to be altered. Severity: high Sources: [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### R-03: Unreviewed code run through a package launcher Local MCP servers are often started by a launcher command such as npx, which fetches whatever the package name resolves to at that moment and runs it. **Impact:** A new publish, a hijacked account or a changed dependency executes on the next start without a review step. **Controls:** - Install a pinned version and run it from a reviewed lockfile or image rather than resolving the latest at launch. - Show the full launch command, unabridged, for approval before a server is added. - Record the package name, version and hash of every approved server in the catalog. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### R-04: Vulnerable or compromised dependency inside a server An MCP server pulls in third-party packages, and a known flaw or a compromised release in any of them runs with the server's permissions. **Impact:** A flaw in a dependency becomes code execution or data access on the machine that runs the server. **Controls:** - Scan server dependencies for known vulnerabilities before approval and on a schedule afterward. - Verify package integrity with checksums or code signing where the publisher provides them. - Subscribe to GitHub advisories for the servers and SDKs in the catalog and track fixed versions. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16). ### R-05: Tool definitions change after approval (rug pull) A server changes its tool names, descriptions or schemas after a user or reviewer approved it, turning a reviewed tool into a different one. **Impact:** An approval given on day one no longer describes what the tool does on day thirty. **Controls:** - Hash the complete tool definition, including inputSchema, outputSchema and annotations, and require re-review when the hash changes. - Re-prompt for consent when definitions change, and alert on listChanged notifications from servers not expected to change. - Pin the server version so definitions only change on a reviewed upgrade. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### R-06: Registry listing mistaken for a security review The official MCP Registry stores metadata that points at packages on npm, PyPI and Docker Hub, and is described as a preview. A listing shows that a name was published, not that the code was reviewed. **Impact:** Teams that build an approved list from registry data may admit a server that nobody read, or map two listings to one entry. The registry's v1.8.8 notes say different server names may share a remote URL. **Controls:** - Treat registry data as discovery input only and review the package behind each entry. - Key catalog entries by server name and version, not by URL. - Record the reviewer, date and version for each approved entry. Severity: medium Sources: [About the MCP Registry](https://modelcontextprotocol.io/registry/about) (read 2026-11-16); [MCP Registry v1.8.8 release notes](https://github.com/modelcontextprotocol/registry/releases/tag/v1.8.8) (read 2026-11-16). ### R-07: Unknown servers configured on developer machines Developers add MCP servers to IDEs, desktop AI apps and command-line agents without telling security. Snyk says its scan data from nearly 10,000 developer environments found 4,524 unique MCP servers in active use. **Impact:** Reviews and policies cover only the servers the team knows about; the rest run with user permissions and no owner. **Controls:** - Inventory MCP servers per machine from the client config files and from an endpoint tool that reports configured servers (Microsoft Defender for Endpoint documents local AI agent discovery that reports each agent's configured MCP servers). - Give every approved server an owner and a review date, and compare the inventory with the catalog on a schedule. - Block or flag servers not in the catalog using managed client settings. Severity: high Sources: [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16); [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint) (read 2026-11-16). ### R-08: Token passthrough and missing audience validation An MCP server accepts tokens that were issued for another service, or forwards the token it received to a downstream API. The MCP spec forbids both. **Impact:** A token stolen from or meant for one service works against another, and downstream logs show the wrong caller. Advisories GHSA-rvq5-wwqv-78pq (TypeScript SDK) and GHSA-w4fh-qvv9-3v23 (Python SDK) describe SDK servers that could not check the audience. **Controls:** - Validate that each access token was issued for this server as its audience (RFC 8707) and reject others. - Never forward the received token to an upstream API; obtain a separate token for downstream calls. - After upgrading an SDK, turn on the audience check the patched version adds, since both advisories say upgrading alone changes nothing. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (read 2026-11-16); [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (read 2026-11-16); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (read 2026-11-16). ### R-09: HTTP endpoint reachable without authentication The protocol makes authorization optional, so a remote or HTTP-exposed MCP server can serve its tools to anyone who reaches it. Advisory CVE-2026-77244 (GHSA-wrhw-j3f9-8vc6) describes an mcp-atlassian token verifier that accepted any non-empty token over HTTP. **Impact:** Anyone with network reach can call the server's tools using the credentials the server holds. **Controls:** - Require authentication on every HTTP endpoint that exposes non-public tools or data. - Run the MCP server's HTTP transport only on loopback or a private network unless it sits behind authentication you control. - Test that a request with an invalid or arbitrary token is refused. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [GHSA-wrhw-j3f9-8vc6: mcp-atlassian authentication bypass (CVE-2026-77244)](https://github.com/advisories/GHSA-wrhw-j3f9-8vc6) (read 2026-11-16). ### R-10: Confused deputy at an MCP proxy server An MCP proxy that fronts a third-party API with one static client ID, allows dynamic client registration and skips per-client consent can be used to obtain authorization codes without the user's approval. **Impact:** An attacker gains access to the third-party API as the user. **Controls:** - Store consent per user and per client_id on the server and check it before starting the upstream flow. - Match redirect URIs by exact string and validate the OAuth state parameter on the callback. - Set consent cookies with the __Host- prefix, Secure, HttpOnly and SameSite=Lax. Severity: medium Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### R-11: Client credentials sent to a server-chosen authorization server An MCP client that trusts server-supplied OAuth metadata can send stored credentials to an authorization server the server names. GHSA-6qxp-vccf-f47h (TypeScript SDK) and GHSA-vqqp-x5qm-5qp4 (FastMCP) describe this class of flaw. **Impact:** A malicious or compromised server can receive a client secret or refresh token with no prompt to the user. **Controls:** - Bind stored credentials to the issuer that granted them and refuse to send them elsewhere (pass expectedIssuer or issuer where the SDK provides it). - Upgrade MCP SDKs to the fixed versions named in the advisories and clear credentials saved without an issuer. - Allow clients to connect only to servers on the approved list. Severity: high Sources: [GHSA-6qxp-vccf-f47h: MCP TypeScript SDK OAuth client credentials (CVE-2026-104850)](https://github.com/advisories/GHSA-6qxp-vccf-f47h) (read 2026-11-16); [GHSA-vqqp-x5qm-5qp4: FastMCP client OAuth issuer binding](https://github.com/advisories/GHSA-vqqp-x5qm-5qp4) (read 2026-11-16). ### R-12: Over-broad token scopes A server requests or holds a wider OAuth scope than its tools need, or all of a user's permissions through a shared token. **Impact:** Any flaw or poisoned instruction in that server acts with the full scope, and one stolen token reaches many resources. **Controls:** - Request the narrowest scopes per tool and use per-server credentials, never one token shared across servers. - Prefer short-lived tokens over long-lived personal access tokens. - Handle insufficient_scope challenges by requesting only the added scope. Severity: medium Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-13: Authorization policy skipped on direct calls A server applies its policy to the original component but a transform, template or direct lookup returns a different one without rechecking. FastMCP advisory GHSA-7xgj-8392-5q4r (23 Oct 2026) describes the case; mcp-atlassian GHSA-3r68-hf9h-887v describes a toolset authorization bypass. **Impact:** A caller who knows an identifier can use a tool or read a resource that the configured policy denies. **Controls:** - Enforce authorization at the handler or server boundary, not only in a layer whose inputs a transform can change. - Write tests that call every protected tool and resource directly, with a low-privilege identity. - Upgrade to the fixed releases the advisories list. Severity: medium Sources: [GHSA-7xgj-8392-5q4r: FastMCP post-transform authorization](https://github.com/advisories/GHSA-7xgj-8392-5q4r) (read 2026-11-16); [GHSA-3r68-hf9h-887v: mcp-atlassian toolset authorization bypass](https://github.com/advisories/GHSA-3r68-hf9h-887v) (read 2026-11-16). ### R-14: Authorization server mix-up and unverified issuer A client talks to more than one authorization server and accepts a response from the wrong one, or accepts metadata whose issuer does not match the server it asked. **Impact:** A code or token meant for one authorization server is redeemed at another. **Controls:** - Record the issuer from validated metadata before redirecting the user and compare it on the response. - Reject authorization server metadata whose issuer is not the expected origin (Python SDK 2.2.0 added this check on the legacy path). - Prefer Client ID Metadata Documents over Dynamic Client Registration, which the 2026-07-28 revision deprecates. Severity: medium Sources: [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### R-15: State handle or message delivered to the wrong client A server treats possession of a state handle as proof of identity, or a shared transport matches messages to clients by request id alone, so one caller's data reaches another. GHSA-jm63-cq47-j5jp describes the second case in the TypeScript server SDK. **Impact:** One user's responses or state are exposed to, or changed by, another user. **Controls:** - Bind handles server-side to the authenticated user and use random, expiring values. - Create the transport or server per request or per session, not once at startup, for HTTP servers. - Verify every inbound request; never treat a handle as authentication. Severity: medium Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [GHSA-jm63-cq47-j5jp: MCP TypeScript server SDK cross-client messages](https://github.com/advisories/GHSA-jm63-cq47-j5jp) (read 2026-11-16). ### R-16: DNS rebinding against a local HTTP server A web page the user visits uses a domain that later resolves to the user's own machine to talk to a local MCP server or developer tool that listens on HTTP. **Impact:** The page can reach tools, tokens or files exposed by the local service. The Inspector's 2.12.0 release says a Host header check closes such a route to its API token and stored OAuth tokens. **Controls:** - Validate the Origin header on every Streamable HTTP request and answer an invalid one with 403, as the spec requires. - Validate the Host header against an allow-list. - Require an authorization token for local HTTP servers, or use stdio. Severity: high Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16). ### R-17: Local server listening on all interfaces A server meant for the local machine binds to 0.0.0.0, so other hosts on the network can connect. **Impact:** Anyone on the network can reach tools that were meant for one user. **Controls:** - Bind to 127.0.0.1 unless network access is intended, as the spec advises. - Use stdio, or a Unix domain socket with restricted permissions, for local servers. - Scan developer subnets for unexpected MCP listeners. Severity: medium Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### R-18: Server-side request forgery through discovery and fetch tools A client follows URLs supplied by a server (OAuth metadata, redirects), or a tool fetches URLs produced by a model, and the target is an internal address or a cloud metadata service. **Impact:** Internal services and cloud credentials become reachable from the machine running the client or server. Inspector 2.15.0 extended its fetch guard to metadata addresses beyond 169.254.0.0/16. **Controls:** - Require HTTPS for OAuth URLs and block private, loopback and link-local ranges with a maintained library, not hand-written parsing. - Validate redirect targets and pin DNS results between check and use. - Send server-side fetches through an egress proxy that enforces destination policy. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Inspector 2.15.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.15.0) (read 2026-11-16). ### R-19: Credentials carried across cross-origin redirects An HTTP client follows a redirect to another origin and sends Authorization headers or cookies with it. Advisories for the Python SDK (GHSA-5h93-6whr-6q8j) and the TypeScript client SDK (GHSA-6prh-2h8m-c8cw) describe the pattern. **Impact:** A server or an intermediary can collect credentials meant for another host. **Controls:** - Follow redirects only within the same scheme, host and port, as Python SDK 2.2.0 does by default. - Drop Authorization and cookie headers on any cross-origin redirect. - Upgrade to the SDK versions named in the advisories. Severity: medium Sources: [GHSA-5h93-6whr-6q8j: MCP Python SDK cross-origin redirects](https://github.com/advisories/GHSA-5h93-6whr-6q8j) (read 2026-11-16); [GHSA-6prh-2h8m-c8cw: MCP TypeScript client SDK cross-origin redirects](https://github.com/advisories/GHSA-6prh-2h8m-c8cw) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### R-20: Remote server without TLS or with unverified identity A client connects to a remote MCP server over plain HTTP, or without checking who the server is. **Impact:** Traffic and tokens can be read or changed in transit, and a different server can answer. **Controls:** - Use TLS for every remote Streamable HTTP connection. - Keep client runtimes current: Inspector 2.13.0 notes that certificate checks rely on the Node.js runtime and names patched Node.js releases after an October 2026 CA issue. - Reject http URLs for OAuth endpoints except loopback addresses during development. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Inspector 2.13.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.13.0) (read 2026-11-16). ### R-21: Unbounded requests and sessions exhaust a server An HTTP server accepts requests or sessions without limits. GHSA-46jr-4wx8-v453 (TypeScript SDK) and GHSA-84m7-p3x7-pcfv (Python SDK) describe memory and session exhaustion. **Impact:** A single unauthenticated client can stop the service. Impact is availability only. **Controls:** - Set request size, session count and idle-timeout limits (Python SDK 2.2.0 expires idle legacy sessions after 30 minutes and caps them at 10,000). - Apply rate limits and quotas per tenant. - Authenticate requests before they reach the SDK. Severity: low Sources: [GHSA-46jr-4wx8-v453: MCP TypeScript SDK memory exhaustion](https://github.com/advisories/GHSA-46jr-4wx8-v453) (read 2026-11-16); [GHSA-84m7-p3x7-pcfv: MCP Python SDK session accumulation](https://github.com/advisories/GHSA-84m7-p3x7-pcfv) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### R-22: Unrestricted outbound access from server processes A server process can open connections to any host, so a flaw or a malicious package can send data anywhere. **Impact:** Stolen data leaves the machine without a block. The altered SP-API package described by Knostic sent credentials to a hardcoded IP. **Controls:** - Disable network access for servers that do not need it and allow-list destinations for those that do. - Log and alert on connections from MCP server processes to new destinations. - Use an egress proxy for server-side deployments. Severity: medium Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16). ### R-23: Poisoned tool description Instructions hidden in a tool description steer the model, which reads the description to decide how to call the tool. Knostic described a package on 12 Nov 2026 whose description tells the agent to fetch a GitHub token without asking the user. **Impact:** The agent takes an action the user did not request, such as reading a credential. **Controls:** - Review the full text of every tool description and parameter field before a server is approved. - Hash and pin the reviewed definitions and alert on change. - Scan descriptions with a tool built for the purpose (the OWASP cheat sheet names mcp-scan) and keep a human approval step for sensitive calls. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token) (read 2026-11-16); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/) (read 2026-11-16). ### R-24: Tool shadowing across servers One server's tool text manipulates how the agent uses tools from another, trusted server, because the model sees all connected descriptions at once. **Impact:** A low-trust server can alter how a high-trust one is used, including which data is sent where. **Controls:** - Treat each server as its own security domain and keep sensitive servers apart from general-purpose ones. - Monitor for cross-server data flows, such as one server's credentials appearing in another's calls. - Limit how many servers a single agent session may load. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-25: Hostile content in schemas and return values Instructions can sit in parameter names, types, defaults and the values a tool returns, not only in the description field. **Impact:** Review that checks only the description misses the other fields the model reads. **Controls:** - Inspect every field of the definition, including inputSchema and outputSchema. - Set additionalProperties to false and use pattern constraints on string fields. - Validate and sanitize tool outputs before they re-enter model context. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### R-26: Annotations trusted as enforcement A client lets a server's readOnlyHint or destructiveHint decide how much approval to ask for. The spec says clients must consider annotations untrusted unless they come from trusted servers. **Impact:** A flipped hint can change approval behavior with no visible change to the description. **Controls:** - Treat annotations as hints only and decide approval from your own policy per tool. - Include annotations in the pinned hash of each tool definition. - Require confirmation for sensitive reads, writes and irreversible actions whatever the label says. Severity: medium Sources: [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-27: Prompt injection through tool results A tool returns text, such as a web page or ticket, that contains instructions the model then follows. **Impact:** An approved server can still cause unapproved actions through the content it returns. **Controls:** - Treat every tool response as untrusted data, including responses from approved servers. - Enforce authorization and validate follow-up tool calls in application code, regardless of the model's interpretation. - Extract structured fields from retrieval tools instead of passing raw pages. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-28: Server-supplied schemas that exhaust or confuse clients A client evaluates patterns or fetches references in a schema a server supplies. Inspector 2.14.0 now checks schema patterns in bounded time, and Python SDK advisory GHSA-rwrf-2pqf-9j8j describes a client fetching server-chosen $ref URLs. **Impact:** A hostile server can slow or break a client, or cause it to make requests it should not. **Controls:** - Evaluate server-supplied patterns with a linear-time matcher and limits. - Do not resolve remote references from schemas supplied by untrusted servers. - Keep MCP clients and the Inspector on current releases. Severity: low Sources: [MCP Inspector 2.14.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.14.0) (read 2026-11-16); [GHSA-rwrf-2pqf-9j8j: MCP Python SDK server-chosen $ref URLs](https://github.com/advisories/GHSA-rwrf-2pqf-9j8j) (read 2026-11-16). ### R-29: Local server runs with the user's full privileges A local MCP server is an ordinary process launched by the host application. It inherits the user's environment variables and can read anything the account can read. **Impact:** A flawed or malicious server reaches SSH keys, cloud credentials, browser profiles and the home directory. **Controls:** - Run local servers in a container or sandbox with minimal default privileges. - Restrict filesystem access to the directories the server needs. - Run servers under a separate account or workspace where the client allows it. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### R-30: Malicious startup command in configuration A client configuration or one-click install carries a launch command that runs code. The spec lists a malicious startup command as a local-server compromise route. **Impact:** Arbitrary code runs with the client's privileges the moment the server starts. **Controls:** - Show the exact command, with arguments and without truncation, before any server is added. - Flag commands that use sudo, delete files, use network tools or reach SSH and home directories. - Distribute approved configurations through managed settings (VS Code allowedMcpServers, Claude Code managed-mcp.json). Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16). ### R-31: Command injection through arguments or URL opening Text from a model, a tool argument or a server-supplied OAuth authorization URL reaches a shell. The Inspector's 2.12.0 notes say a remote server's authorization URL could previously run commands through the opener on Windows and WSL. **Impact:** Code runs on the machine with the user's privileges. **Controls:** - Allow only http and https schemes for authorization URLs and never open URLs through a shell. - Never pass raw shell commands or unsanitized paths from tool arguments. - Validate tool inputs against strict schemas and allow-lists. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-32: Path traversal and arbitrary file read in file tools A tool that reads or uploads files accepts a path without limiting it to an intended directory. 14 of the 25 mcp-atlassian advisories published on 22 Sep 2026 concern file read through attachment tools. **Impact:** Server credentials and local files can be read and sent out through the tool. **Controls:** - Resolve paths on the server and reject any that leave an allowed directory. - Run file tools in a sandbox that mounts only the needed folders. - Upgrade mcp-atlassian to 0.22.0 or later. Severity: high Sources: [GHSA-93xw-j965-9mx3: mcp-atlassian arbitrary file read](https://github.com/advisories/GHSA-93xw-j965-9mx3) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-33: Project configuration from a cloned repository A repository carries MCP configuration that a client could load automatically. Claude Code's docs say a cloned repository cannot approve its own servers and that approval settings committed to the project are ignored in an untrusted folder. **Impact:** Opening a repository could start a server chosen by its author. **Controls:** - Do not auto-enable project servers: leave enableAllProjectMcpServers unset or false and review each project server. - Use workspace trust and require approval for project-scoped servers. - Lock server sets with managed-mcp.json and allowedMcpServers or deniedMcpServers. Severity: high Sources: [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16); [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings) (read 2026-11-16). ### R-34: Tool calls auto-approved without full arguments A client approves tool calls automatically, or shows only a tool name, so the user never sees what is about to run. **Impact:** A poisoned description or injected content can trigger destructive or data-sharing actions without anyone noticing. **Controls:** - Require explicit confirmation for destructive, financial and data-sharing operations and show full parameters. - Never auto-approve tool calls in multi-server setups. - Make sure the confirmation interface cannot be satisfied by text the model generates. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### R-35: Developer tools that spawn processes on a port A developer tool such as an MCP inspector can start stdio servers on request, and its API token is what protects that ability from other callers. **Impact:** A weak guard turns a testing tool into a way to run commands. The Inspector's own docs say its web backend can spawn processes and that stdio servers run as the user. **Controls:** - Keep the Inspector on a current version; releases 2.12.0, 2.14.0 and 2.15.0 carry security fixes. - Do not disable its authentication (DANGEROUSLY_OMIT_AUTH) outside an isolated test setup. - Bind it to loopback and never expose its port. Severity: medium Sources: [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16). ### R-36: Secrets inherited through the environment A server started by the client inherits the environment of the host process, which often holds cloud keys and tokens. **Impact:** A server that needs one key can read all of them. **Controls:** - Start servers with a minimal environment and pass only the variables each one needs. - Use separate credentials per server with the least privilege. - Keep secrets in an OS credential store rather than in the shell profile. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-37: Tokens stored in plaintext A server or client saves OAuth tokens in a file with loose permissions or in plain text. mcp-atlassian advisories GHSA-g5xv-mhgm-v5f6 and GHSA-4596-2p6p-28cv describe this, and the Inspector's docs say its secrets fall back to a plaintext file when no OS keychain exists. **Impact:** Any local process or user that can read the file gets the tokens. **Controls:** - Use the OS credential store (macOS Keychain, Windows Credential Manager, Linux Secret Service). - Set file permissions to the owner only where files are unavoidable. - Never store tokens in MCP config files, which the OWASP cheat sheet also says. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [GHSA-g5xv-mhgm-v5f6: mcp-atlassian plaintext OAuth token storage](https://github.com/advisories/GHSA-g5xv-mhgm-v5f6) (read 2026-11-16); [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (read 2026-11-16). ### R-38: Data exfiltration through legitimate tool calls Injected instructions encode sensitive data into ordinary-looking calls, such as search queries or email subjects. **Impact:** Data leaves through channels that policy allows. **Controls:** - Monitor for credentials or sensitive values from one server appearing in calls to another. - Restrict which tools can send data outside the organization and require approval for them. - Limit the data each server can read in the first place. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-39: Sensitive data requested through form elicitation A server uses form-mode elicitation to ask the user for a password, API key or token, which passes through the client. **Impact:** Secrets pass through the client and its logs. The spec forbids it. **Controls:** - Reject servers that request secrets in form mode and use URL mode for credential steps. - Show which server is asking and let the user decline. - Review elicitation requests during server vetting. Severity: medium Sources: [MCP Elicitation, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/client/elicitation) (read 2026-11-16). ### R-40: No record of tool calls Tool invocations are not logged with their parameters, user and time, so there is nothing to investigate after an incident. **Impact:** A misuse or a poisoned call cannot be reconstructed, and anomalies go unseen. **Controls:** - Log every tool invocation with full parameters, user context and timestamp. - Send the logs to a SIEM and alert on new tools, admin-level queries and unusual call rates. - Keep the logs long enough to cover an investigation window you set. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### R-41: Secrets and personal data written into logs Logging full parameters also records secrets and personal data that appear in arguments and results. **Impact:** The log store becomes a second place to steal credentials from. **Controls:** - Redact secrets and personal data before storing tool-call logs. - Restrict who can read the logs and set a retention period. - Test the redaction rules with sample credentials. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-42: Protected data returned through resources and templates Resources and resource templates expose data by URI, and a template can retrieve a resource that a rule denied. **Impact:** Data outside a caller's rights is returned. FastMCP's GHSA-7xgj fix makes a denied concrete resource shadow a matching template. **Controls:** - Check authorization for the final resource that is returned, not the template. - Avoid overlapping proxy templates for protected resources. - Test reads for denied URIs through every route. Severity: medium Sources: [GHSA-7xgj-8392-5q4r: FastMCP post-transform authorization](https://github.com/advisories/GHSA-7xgj-8392-5q4r) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ## Download and cite [Download the MCP Risk Register as CSV or JSON](/datasets/mcp-risk-register/). Cite it as: Cite assets as: MCP Server Security, "", edition 16 Nov 2026, https://mcpservercybersecurity.com//, licensed CC BY 4.0. ## Related pages - [How the MCP Risk Register is built](/mcp-risk-register/methodology/) - [MCP Risk Register data download](/datasets/mcp-risk-register/) - [Supply chain](/mcp-risk-register/supply-chain/) - [Authorization](/mcp-risk-register/authorization/) - [Transport and network](/mcp-risk-register/transport-and-network/) - [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) - [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) - [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) - [MCP Security Checklist](/mcp-security-checklist/) - [The ranking of tools](/ranking/) --- # How the MCP Risk Register is built URL: https://mcpservercybersecurity.com/mcp-risk-register/methodology/ Updated: 2026-11-16 The register lists 42 risks across six dimensions for MCP servers. Every row carries what the risk is, its impact, controls, a severity set by one written rule, and primary sources read on 16 Nov 2026. It is released under CC BY 4.0, and the sections below say how to cite it and report corrections. ## What the register is The register is a list of 42 risks that affect MCP servers and the clients that run them, grouped into six dimensions: supply chain, authorization, transport and network, tool descriptions and schemas, execution and filesystem, and data exposure and logging. Rows are written for a platform or security engineer whose developers install MCP servers. A row describes a risk class and its fixes. It carries no exploit code, payloads or bypass steps. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources and read date Every row cites at least one primary source: the MCP specification and its security pages, the OWASP MCP Security Cheat Sheet, GitHub security advisories and CVE records, release notes from the maintainers, or research published by the named lab or vendor. All pages were fetched and read on 16 Nov 2026 (read date 2026-11-16), and only pages that returned HTTP 200 are cited. Vendor research is described as the vendor's own account. The 2026-07-28 MCP revision is the reference for spec statements; where an older revision differs, the row says so. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization); [OWASP MCP Top 10](https://owasp.org/www-project-mcp-top-10/). ## Severity rule Severity follows one rule applied to every row. High: the risk can lead to code execution on a machine, theft of a credential, or access to data or tools without authentication, and it needs no more than normal use of the server (installing it, connecting to it, visiting a web page, or leaving it reachable). Medium: the same kinds of impact, but a precondition is needed, such as a particular configuration, a specific transport, an authenticated caller or a second flaw. Low: the impact is loss of availability or a weakness that matters only alongside another risk. Applied on 16 Nov 2026 the register has 19 high, 21 medium and 2 low rows. The rule is about the class of risk, not about any single product, and a given deployment may be more or less exposed. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices). ## How to cite Cite as: MCP Risk Register, MCP Server Security, edition of 16 Nov 2026, https://mcpservercybersecurity.com/mcp-risk-register/, licensed CC BY 4.0. Cite a row by its identifier (for example R-08) and the edition date. ## License The register text and the downloadable CSV and JSON files are released under Creative Commons Attribution 4.0 International (CC BY 4.0). You may copy and adapt them with credit and a link to the license. ## Corrections If a source has changed, a row is wrong or a risk is missing, use the corrections page on this site. Corrections are dated and listed there, and the register edition date changes when rows change. ## Counts - 42 items in 6 categories. - 19 with severity high. - 21 with severity medium. - 2 with severity low. ## Related pages - [The MCP Risk Register](/mcp-risk-register/) - [MCP Risk Register download](/datasets/mcp-risk-register/) - [Scoring method](/method/) - [Supply chain](/mcp-risk-register/supply-chain/) - [Authorization](/mcp-risk-register/authorization/) - [Transport and network](/mcp-risk-register/transport-and-network/) --- # MCP supply chain risks URL: https://mcpservercybersecurity.com/mcp-risk-register/supply-chain/ Updated: 2026-11-16 Seven supply chain risks cover how an MCP server reaches a machine: look-alike names, repackaged code, launcher commands, dependencies, changed tool definitions, registry listings and servers nobody inventoried. Each row names a control and a primary source, with severity set by the written rule, as read on 16 Nov 2026. ## What this dimension covers How a server reaches the machine: names, packages, launchers, dependencies, registries and unknown installs. The register holds 7 rows here (R-01, R-02, R-03, R-04, R-05, R-06, R-07). Each row states what the risk is, its impact, concrete controls and at least one primary source. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip). ## Highest-severity rows Rows that meet the high rule under the written rule are: R-01 Look-alike package name (typosquatting); R-02 Repackaged server with altered endpoints; R-03 Unreviewed code run through a package launcher; R-07 Unknown servers configured on developer machines. - R-01: The wrong package runs with the same permissions as the right one, including access to the user's files and environment variables. - R-02: Application credentials the user supplies as environment variables go to a destination the user did not choose. - R-03: A new publish, a hijacked account or a changed dependency executes on the next start without a review step. - R-07: Reviews and policies cover only the servers the team knows about; the rest run with user permissions and no owner. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices). ## First controls to apply Start with the controls on the high rows, then work down. Controls name settings only where a primary source documents them; check each against the version your clients run. - R-01: Compare the exact package name, publisher and repository against the vendor's documentation before the first install. - R-02: Diff the package contents against the upstream repository for endpoint and authentication changes before approval. - R-03: Install a pinned version and run it from a reviewed lockfile or image rather than resolving the latest at launch. - R-04: Scan server dependencies for known vulnerabilities before approval and on a schedule afterward. - R-05: Hash the complete tool definition, including inputSchema, outputSchema and annotations, and require re-review when the hash changes. - R-06: Treat registry data as discovery input only and review the package behind each entry. - R-07: Inventory MCP servers per machine from the client config files and from an endpoint tool that reports configured servers (Microsoft Defender for Endpoint documents local AI agent discovery that reports each agent's configured MCP servers). Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices). ## Supply chain: 7 items ### R-01: Look-alike package name (typosquatting) A package is published under a name one keystroke away from a trusted MCP server, in the hope that a developer or a script installs it by mistake. **Impact:** The wrong package runs with the same permissions as the right one, including access to the user's files and environment variables. **Controls:** - Compare the exact package name, publisher and repository against the vendor's documentation before the first install. - Keep an approved server list and block installs of anything not on it (VS Code and Copilot enterprise-managed settings use allowedMcpServers and deniedMcpServers; Claude Code uses managed-mcp.json). - Pin the version once reviewed so a later publish under the same name is not picked up. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### R-02: Repackaged server with altered endpoints A package copies a legitimate server's open-source code but replaces authentication or API addresses, and the README does not mention the change. Knostic described one such npm package on 26 Oct 2026 and assessed it as suspected malicious credential redirection. **Impact:** Application credentials the user supplies as environment variables go to a destination the user did not choose. **Controls:** - Diff the package contents against the upstream repository for endpoint and authentication changes before approval. - Restrict outbound network access from MCP server processes to the hosts the server needs. - Rotate any credential that was supplied to a server later found to be altered. Severity: high Sources: [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### R-03: Unreviewed code run through a package launcher Local MCP servers are often started by a launcher command such as npx, which fetches whatever the package name resolves to at that moment and runs it. **Impact:** A new publish, a hijacked account or a changed dependency executes on the next start without a review step. **Controls:** - Install a pinned version and run it from a reviewed lockfile or image rather than resolving the latest at launch. - Show the full launch command, unabridged, for approval before a server is added. - Record the package name, version and hash of every approved server in the catalog. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### R-04: Vulnerable or compromised dependency inside a server An MCP server pulls in third-party packages, and a known flaw or a compromised release in any of them runs with the server's permissions. **Impact:** A flaw in a dependency becomes code execution or data access on the machine that runs the server. **Controls:** - Scan server dependencies for known vulnerabilities before approval and on a schedule afterward. - Verify package integrity with checksums or code signing where the publisher provides them. - Subscribe to GitHub advisories for the servers and SDKs in the catalog and track fixed versions. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16). ### R-05: Tool definitions change after approval (rug pull) A server changes its tool names, descriptions or schemas after a user or reviewer approved it, turning a reviewed tool into a different one. **Impact:** An approval given on day one no longer describes what the tool does on day thirty. **Controls:** - Hash the complete tool definition, including inputSchema, outputSchema and annotations, and require re-review when the hash changes. - Re-prompt for consent when definitions change, and alert on listChanged notifications from servers not expected to change. - Pin the server version so definitions only change on a reviewed upgrade. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### R-06: Registry listing mistaken for a security review The official MCP Registry stores metadata that points at packages on npm, PyPI and Docker Hub, and is described as a preview. A listing shows that a name was published, not that the code was reviewed. **Impact:** Teams that build an approved list from registry data may admit a server that nobody read, or map two listings to one entry. The registry's v1.8.8 notes say different server names may share a remote URL. **Controls:** - Treat registry data as discovery input only and review the package behind each entry. - Key catalog entries by server name and version, not by URL. - Record the reviewer, date and version for each approved entry. Severity: medium Sources: [About the MCP Registry](https://modelcontextprotocol.io/registry/about) (read 2026-11-16); [MCP Registry v1.8.8 release notes](https://github.com/modelcontextprotocol/registry/releases/tag/v1.8.8) (read 2026-11-16). ### R-07: Unknown servers configured on developer machines Developers add MCP servers to IDEs, desktop AI apps and command-line agents without telling security. Snyk says its scan data from nearly 10,000 developer environments found 4,524 unique MCP servers in active use. **Impact:** Reviews and policies cover only the servers the team knows about; the rest run with user permissions and no owner. **Controls:** - Inventory MCP servers per machine from the client config files and from an endpoint tool that reports configured servers (Microsoft Defender for Endpoint documents local AI agent discovery that reports each agent's configured MCP servers). - Give every approved server an owner and a review date, and compare the inventory with the catalog on a schedule. - Block or flag servers not in the catalog using managed client settings. Severity: high Sources: [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16); [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint) (read 2026-11-16). ## Related pages - [The full MCP Risk Register](/mcp-risk-register/) - [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) - [Authorization](/mcp-risk-register/authorization/) - [Method](/mcp-risk-register/methodology/) - [MCP Security Checklist](/mcp-security-checklist/) --- # MCP authorization risks URL: https://mcpservercybersecurity.com/mcp-risk-register/authorization/ Updated: 2026-11-16 Eight authorization risks cover token audience, unauthenticated HTTP endpoints, proxy consent, credentials sent to the wrong issuer, broad scopes, skipped policy checks, issuer mix-up and cross-client state. Each row cites the MCP authorization spec, an advisory or both, with severity set by one written rule, read on 16 Nov 2026. ## What this dimension covers Who may call a server and what a token allows: audience, scopes, issuers, consent and state. The register holds 8 rows here (R-08, R-09, R-10, R-11, R-12, R-13, R-14, R-15). Each row states what the risk is, its impact, concrete controls and at least one primary source. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization); [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq). ## Highest-severity rows Rows that meet the high rule under the written rule are: R-08 Token passthrough and missing audience validation; R-09 HTTP endpoint reachable without authentication; R-11 Client credentials sent to a server-chosen authorization server. - R-08: A token stolen from or meant for one service works against another, and downstream logs show the wrong caller. Advisories GHSA-rvq5-wwqv-78pq (TypeScript SDK) and GHSA-w4fh-qvv9-3v23 (Python SDK) describe SDK servers that could not check the audience. - R-09: Anyone with network reach can call the server's tools using the credentials the server holds. - R-11: A malicious or compromised server can receive a client secret or refresh token with no prompt to the user. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization); [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23). ## First controls to apply Start with the controls on the high rows, then work down. Controls name settings only where a primary source documents them; check each against the version your clients run. - R-08: Validate that each access token was issued for this server as its audience (RFC 8707) and reject others. - R-09: Require authentication on every HTTP endpoint that exposes non-public tools or data. - R-10: Store consent per user and per client_id on the server and check it before starting the upstream flow. - R-11: Bind stored credentials to the issuer that granted them and refuse to send them elsewhere (pass expectedIssuer or issuer where the SDK provides it). - R-12: Request the narrowest scopes per tool and use per-server credentials, never one token shared across servers. - R-13: Enforce authorization at the handler or server boundary, not only in a layer whose inputs a transform can change. - R-14: Record the issuer from validated metadata before redirecting the user and compare it on the response. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization); [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23). ## Authorization: 8 items ### R-08: Token passthrough and missing audience validation An MCP server accepts tokens that were issued for another service, or forwards the token it received to a downstream API. The MCP spec forbids both. **Impact:** A token stolen from or meant for one service works against another, and downstream logs show the wrong caller. Advisories GHSA-rvq5-wwqv-78pq (TypeScript SDK) and GHSA-w4fh-qvv9-3v23 (Python SDK) describe SDK servers that could not check the audience. **Controls:** - Validate that each access token was issued for this server as its audience (RFC 8707) and reject others. - Never forward the received token to an upstream API; obtain a separate token for downstream calls. - After upgrading an SDK, turn on the audience check the patched version adds, since both advisories say upgrading alone changes nothing. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (read 2026-11-16); [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (read 2026-11-16); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (read 2026-11-16). ### R-09: HTTP endpoint reachable without authentication The protocol makes authorization optional, so a remote or HTTP-exposed MCP server can serve its tools to anyone who reaches it. Advisory CVE-2026-77244 (GHSA-wrhw-j3f9-8vc6) describes an mcp-atlassian token verifier that accepted any non-empty token over HTTP. **Impact:** Anyone with network reach can call the server's tools using the credentials the server holds. **Controls:** - Require authentication on every HTTP endpoint that exposes non-public tools or data. - Run the MCP server's HTTP transport only on loopback or a private network unless it sits behind authentication you control. - Test that a request with an invalid or arbitrary token is refused. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [GHSA-wrhw-j3f9-8vc6: mcp-atlassian authentication bypass (CVE-2026-77244)](https://github.com/advisories/GHSA-wrhw-j3f9-8vc6) (read 2026-11-16). ### R-10: Confused deputy at an MCP proxy server An MCP proxy that fronts a third-party API with one static client ID, allows dynamic client registration and skips per-client consent can be used to obtain authorization codes without the user's approval. **Impact:** An attacker gains access to the third-party API as the user. **Controls:** - Store consent per user and per client_id on the server and check it before starting the upstream flow. - Match redirect URIs by exact string and validate the OAuth state parameter on the callback. - Set consent cookies with the __Host- prefix, Secure, HttpOnly and SameSite=Lax. Severity: medium Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### R-11: Client credentials sent to a server-chosen authorization server An MCP client that trusts server-supplied OAuth metadata can send stored credentials to an authorization server the server names. GHSA-6qxp-vccf-f47h (TypeScript SDK) and GHSA-vqqp-x5qm-5qp4 (FastMCP) describe this class of flaw. **Impact:** A malicious or compromised server can receive a client secret or refresh token with no prompt to the user. **Controls:** - Bind stored credentials to the issuer that granted them and refuse to send them elsewhere (pass expectedIssuer or issuer where the SDK provides it). - Upgrade MCP SDKs to the fixed versions named in the advisories and clear credentials saved without an issuer. - Allow clients to connect only to servers on the approved list. Severity: high Sources: [GHSA-6qxp-vccf-f47h: MCP TypeScript SDK OAuth client credentials (CVE-2026-104850)](https://github.com/advisories/GHSA-6qxp-vccf-f47h) (read 2026-11-16); [GHSA-vqqp-x5qm-5qp4: FastMCP client OAuth issuer binding](https://github.com/advisories/GHSA-vqqp-x5qm-5qp4) (read 2026-11-16). ### R-12: Over-broad token scopes A server requests or holds a wider OAuth scope than its tools need, or all of a user's permissions through a shared token. **Impact:** Any flaw or poisoned instruction in that server acts with the full scope, and one stolen token reaches many resources. **Controls:** - Request the narrowest scopes per tool and use per-server credentials, never one token shared across servers. - Prefer short-lived tokens over long-lived personal access tokens. - Handle insufficient_scope challenges by requesting only the added scope. Severity: medium Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-13: Authorization policy skipped on direct calls A server applies its policy to the original component but a transform, template or direct lookup returns a different one without rechecking. FastMCP advisory GHSA-7xgj-8392-5q4r (23 Oct 2026) describes the case; mcp-atlassian GHSA-3r68-hf9h-887v describes a toolset authorization bypass. **Impact:** A caller who knows an identifier can use a tool or read a resource that the configured policy denies. **Controls:** - Enforce authorization at the handler or server boundary, not only in a layer whose inputs a transform can change. - Write tests that call every protected tool and resource directly, with a low-privilege identity. - Upgrade to the fixed releases the advisories list. Severity: medium Sources: [GHSA-7xgj-8392-5q4r: FastMCP post-transform authorization](https://github.com/advisories/GHSA-7xgj-8392-5q4r) (read 2026-11-16); [GHSA-3r68-hf9h-887v: mcp-atlassian toolset authorization bypass](https://github.com/advisories/GHSA-3r68-hf9h-887v) (read 2026-11-16). ### R-14: Authorization server mix-up and unverified issuer A client talks to more than one authorization server and accepts a response from the wrong one, or accepts metadata whose issuer does not match the server it asked. **Impact:** A code or token meant for one authorization server is redeemed at another. **Controls:** - Record the issuer from validated metadata before redirecting the user and compare it on the response. - Reject authorization server metadata whose issuer is not the expected origin (Python SDK 2.2.0 added this check on the legacy path). - Prefer Client ID Metadata Documents over Dynamic Client Registration, which the 2026-07-28 revision deprecates. Severity: medium Sources: [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### R-15: State handle or message delivered to the wrong client A server treats possession of a state handle as proof of identity, or a shared transport matches messages to clients by request id alone, so one caller's data reaches another. GHSA-jm63-cq47-j5jp describes the second case in the TypeScript server SDK. **Impact:** One user's responses or state are exposed to, or changed by, another user. **Controls:** - Bind handles server-side to the authenticated user and use random, expiring values. - Create the transport or server per request or per session, not once at startup, for HTTP servers. - Verify every inbound request; never treat a handle as authentication. Severity: medium Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [GHSA-jm63-cq47-j5jp: MCP TypeScript server SDK cross-client messages](https://github.com/advisories/GHSA-jm63-cq47-j5jp) (read 2026-11-16). ## Related pages - [The full MCP Risk Register](/mcp-risk-register/) - [Supply chain](/mcp-risk-register/supply-chain/) - [Transport and network](/mcp-risk-register/transport-and-network/) - [Method](/mcp-risk-register/methodology/) - [MCP Security Checklist](/mcp-security-checklist/) --- # MCP transport and network risks URL: https://mcpservercybersecurity.com/mcp-risk-register/transport-and-network/ Updated: 2026-11-16 Seven transport and network risks cover DNS rebinding, listeners on all interfaces, server-side request forgery, cross-origin redirects, missing TLS, unbounded requests and open egress. Each row gives a control and a primary source from the MCP spec, release notes or an advisory, with severity set by the written rule, read on 16 Nov 2026. ## What this dimension covers The channel between client and server: local listeners, TLS, redirects, request fetching and egress. The register holds 7 rows here (R-16, R-17, R-18, R-19, R-20, R-21, R-22). Each row states what the risk is, its impact, concrete controls and at least one primary source. Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices). ## Highest-severity rows Rows that meet the high rule under the written rule are: R-16 DNS rebinding against a local HTTP server; R-18 Server-side request forgery through discovery and fetch tools. - R-16: The page can reach tools, tokens or files exposed by the local service. The Inspector's 2.12.0 release says a Host header check closes such a route to its API token and stored OAuth tokens. - R-18: Internal services and cloud credentials become reachable from the machine running the client or server. Inspector 2.15.0 extended its fetch guard to metadata addresses beyond 169.254.0.0/16. Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Inspector 2.15.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.15.0). ## First controls to apply Start with the controls on the high rows, then work down. Controls name settings only where a primary source documents them; check each against the version your clients run. - R-16: Validate the Origin header on every Streamable HTTP request and answer an invalid one with 403, as the spec requires. - R-17: Bind to 127.0.0.1 unless network access is intended, as the spec advises. - R-18: Require HTTPS for OAuth URLs and block private, loopback and link-local ranges with a maintained library, not hand-written parsing. - R-19: Follow redirects only within the same scheme, host and port, as Python SDK 2.2.0 does by default. - R-20: Use TLS for every remote Streamable HTTP connection. - R-21: Set request size, session count and idle-timeout limits (Python SDK 2.2.0 expires idle legacy sessions after 30 minutes and caps them at 10,000). - R-22: Disable network access for servers that do not need it and allow-list destinations for those that do. Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Inspector 2.15.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.15.0). ## Transport and network: 7 items ### R-16: DNS rebinding against a local HTTP server A web page the user visits uses a domain that later resolves to the user's own machine to talk to a local MCP server or developer tool that listens on HTTP. **Impact:** The page can reach tools, tokens or files exposed by the local service. The Inspector's 2.12.0 release says a Host header check closes such a route to its API token and stored OAuth tokens. **Controls:** - Validate the Origin header on every Streamable HTTP request and answer an invalid one with 403, as the spec requires. - Validate the Host header against an allow-list. - Require an authorization token for local HTTP servers, or use stdio. Severity: high Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16). ### R-17: Local server listening on all interfaces A server meant for the local machine binds to 0.0.0.0, so other hosts on the network can connect. **Impact:** Anyone on the network can reach tools that were meant for one user. **Controls:** - Bind to 127.0.0.1 unless network access is intended, as the spec advises. - Use stdio, or a Unix domain socket with restricted permissions, for local servers. - Scan developer subnets for unexpected MCP listeners. Severity: medium Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### R-18: Server-side request forgery through discovery and fetch tools A client follows URLs supplied by a server (OAuth metadata, redirects), or a tool fetches URLs produced by a model, and the target is an internal address or a cloud metadata service. **Impact:** Internal services and cloud credentials become reachable from the machine running the client or server. Inspector 2.15.0 extended its fetch guard to metadata addresses beyond 169.254.0.0/16. **Controls:** - Require HTTPS for OAuth URLs and block private, loopback and link-local ranges with a maintained library, not hand-written parsing. - Validate redirect targets and pin DNS results between check and use. - Send server-side fetches through an egress proxy that enforces destination policy. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Inspector 2.15.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.15.0) (read 2026-11-16). ### R-19: Credentials carried across cross-origin redirects An HTTP client follows a redirect to another origin and sends Authorization headers or cookies with it. Advisories for the Python SDK (GHSA-5h93-6whr-6q8j) and the TypeScript client SDK (GHSA-6prh-2h8m-c8cw) describe the pattern. **Impact:** A server or an intermediary can collect credentials meant for another host. **Controls:** - Follow redirects only within the same scheme, host and port, as Python SDK 2.2.0 does by default. - Drop Authorization and cookie headers on any cross-origin redirect. - Upgrade to the SDK versions named in the advisories. Severity: medium Sources: [GHSA-5h93-6whr-6q8j: MCP Python SDK cross-origin redirects](https://github.com/advisories/GHSA-5h93-6whr-6q8j) (read 2026-11-16); [GHSA-6prh-2h8m-c8cw: MCP TypeScript client SDK cross-origin redirects](https://github.com/advisories/GHSA-6prh-2h8m-c8cw) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### R-20: Remote server without TLS or with unverified identity A client connects to a remote MCP server over plain HTTP, or without checking who the server is. **Impact:** Traffic and tokens can be read or changed in transit, and a different server can answer. **Controls:** - Use TLS for every remote Streamable HTTP connection. - Keep client runtimes current: Inspector 2.13.0 notes that certificate checks rely on the Node.js runtime and names patched Node.js releases after an October 2026 CA issue. - Reject http URLs for OAuth endpoints except loopback addresses during development. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Inspector 2.13.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.13.0) (read 2026-11-16). ### R-21: Unbounded requests and sessions exhaust a server An HTTP server accepts requests or sessions without limits. GHSA-46jr-4wx8-v453 (TypeScript SDK) and GHSA-84m7-p3x7-pcfv (Python SDK) describe memory and session exhaustion. **Impact:** A single unauthenticated client can stop the service. Impact is availability only. **Controls:** - Set request size, session count and idle-timeout limits (Python SDK 2.2.0 expires idle legacy sessions after 30 minutes and caps them at 10,000). - Apply rate limits and quotas per tenant. - Authenticate requests before they reach the SDK. Severity: low Sources: [GHSA-46jr-4wx8-v453: MCP TypeScript SDK memory exhaustion](https://github.com/advisories/GHSA-46jr-4wx8-v453) (read 2026-11-16); [GHSA-84m7-p3x7-pcfv: MCP Python SDK session accumulation](https://github.com/advisories/GHSA-84m7-p3x7-pcfv) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### R-22: Unrestricted outbound access from server processes A server process can open connections to any host, so a flaw or a malicious package can send data anywhere. **Impact:** Stolen data leaves the machine without a block. The altered SP-API package described by Knostic sent credentials to a hardcoded IP. **Controls:** - Disable network access for servers that do not need it and allow-list destinations for those that do. - Log and alert on connections from MCP server processes to new destinations. - Use an egress proxy for server-side deployments. Severity: medium Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16). ## Related pages - [The full MCP Risk Register](/mcp-risk-register/) - [Authorization](/mcp-risk-register/authorization/) - [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) - [Method](/mcp-risk-register/methodology/) - [MCP Security Checklist](/mcp-security-checklist/) --- # MCP tool descriptions and schemas risks URL: https://mcpservercybersecurity.com/mcp-risk-register/tool-descriptions-and-schemas/ Updated: 2026-11-16 Six risks sit in what the model reads: poisoned descriptions, shadowing across servers, hostile schemas and return values, trusted annotations, prompt injection in tool results and server-supplied schemas that burden clients. Each row cites OWASP, the MCP spec or published research, with severity from one written rule, read on 16 Nov 2026. ## What this dimension covers What the model reads: descriptions, schemas, annotations and tool results. The register holds 6 rows here (R-23, R-24, R-25, R-26, R-27, R-28). Each row states what the risk is, its impact, concrete controls and at least one primary source. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/). ## Highest-severity rows Rows that meet the high rule under the written rule are: R-23 Poisoned tool description; R-27 Prompt injection through tool results. - R-23: The agent takes an action the user did not request, such as reading a credential. - R-27: An approved server can still cause unapproved actions through the content it returns. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools). ## First controls to apply Start with the controls on the high rows, then work down. Controls name settings only where a primary source documents them; check each against the version your clients run. - R-23: Review the full text of every tool description and parameter field before a server is approved. - R-24: Treat each server as its own security domain and keep sensitive servers apart from general-purpose ones. - R-25: Inspect every field of the definition, including inputSchema and outputSchema. - R-26: Treat annotations as hints only and decide approval from your own policy per tool. - R-27: Treat every tool response as untrusted data, including responses from approved servers. - R-28: Evaluate server-supplied patterns with a linear-time matcher and limits. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools). ## Tool descriptions and schemas: 6 items ### R-23: Poisoned tool description Instructions hidden in a tool description steer the model, which reads the description to decide how to call the tool. Knostic described a package on 12 Nov 2026 whose description tells the agent to fetch a GitHub token without asking the user. **Impact:** The agent takes an action the user did not request, such as reading a credential. **Controls:** - Review the full text of every tool description and parameter field before a server is approved. - Hash and pin the reviewed definitions and alert on change. - Scan descriptions with a tool built for the purpose (the OWASP cheat sheet names mcp-scan) and keep a human approval step for sensitive calls. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token) (read 2026-11-16); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/) (read 2026-11-16). ### R-24: Tool shadowing across servers One server's tool text manipulates how the agent uses tools from another, trusted server, because the model sees all connected descriptions at once. **Impact:** A low-trust server can alter how a high-trust one is used, including which data is sent where. **Controls:** - Treat each server as its own security domain and keep sensitive servers apart from general-purpose ones. - Monitor for cross-server data flows, such as one server's credentials appearing in another's calls. - Limit how many servers a single agent session may load. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-25: Hostile content in schemas and return values Instructions can sit in parameter names, types, defaults and the values a tool returns, not only in the description field. **Impact:** Review that checks only the description misses the other fields the model reads. **Controls:** - Inspect every field of the definition, including inputSchema and outputSchema. - Set additionalProperties to false and use pattern constraints on string fields. - Validate and sanitize tool outputs before they re-enter model context. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### R-26: Annotations trusted as enforcement A client lets a server's readOnlyHint or destructiveHint decide how much approval to ask for. The spec says clients must consider annotations untrusted unless they come from trusted servers. **Impact:** A flipped hint can change approval behavior with no visible change to the description. **Controls:** - Treat annotations as hints only and decide approval from your own policy per tool. - Include annotations in the pinned hash of each tool definition. - Require confirmation for sensitive reads, writes and irreversible actions whatever the label says. Severity: medium Sources: [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-27: Prompt injection through tool results A tool returns text, such as a web page or ticket, that contains instructions the model then follows. **Impact:** An approved server can still cause unapproved actions through the content it returns. **Controls:** - Treat every tool response as untrusted data, including responses from approved servers. - Enforce authorization and validate follow-up tool calls in application code, regardless of the model's interpretation. - Extract structured fields from retrieval tools instead of passing raw pages. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-28: Server-supplied schemas that exhaust or confuse clients A client evaluates patterns or fetches references in a schema a server supplies. Inspector 2.14.0 now checks schema patterns in bounded time, and Python SDK advisory GHSA-rwrf-2pqf-9j8j describes a client fetching server-chosen $ref URLs. **Impact:** A hostile server can slow or break a client, or cause it to make requests it should not. **Controls:** - Evaluate server-supplied patterns with a linear-time matcher and limits. - Do not resolve remote references from schemas supplied by untrusted servers. - Keep MCP clients and the Inspector on current releases. Severity: low Sources: [MCP Inspector 2.14.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.14.0) (read 2026-11-16); [GHSA-rwrf-2pqf-9j8j: MCP Python SDK server-chosen $ref URLs](https://github.com/advisories/GHSA-rwrf-2pqf-9j8j) (read 2026-11-16). ## Related pages - [The full MCP Risk Register](/mcp-risk-register/) - [Transport and network](/mcp-risk-register/transport-and-network/) - [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) - [Method](/mcp-risk-register/methodology/) - [MCP Security Checklist](/mcp-security-checklist/) --- # MCP execution and filesystem risks URL: https://mcpservercybersecurity.com/mcp-risk-register/execution-and-filesystem/ Updated: 2026-11-16 Seven risks cover what a server can run and reach on the host: full user privileges, malicious startup commands, command injection, path traversal, project configuration from cloned repositories, auto-approved tool calls and developer tools that spawn processes. Each row has controls and a primary source, with severity from one written rule, read on 16 Nov 2026. ## What this dimension covers What a server can run and reach on the host: privileges, commands, paths and approvals. The register holds 7 rows here (R-29, R-30, R-31, R-32, R-33, R-34, R-35). Each row states what the risk is, its impact, concrete controls and at least one primary source. Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp). ## Highest-severity rows Rows that meet the high rule under the written rule are: R-29 Local server runs with the user's full privileges; R-30 Malicious startup command in configuration; R-31 Command injection through arguments or URL opening; R-32 Path traversal and arbitrary file read in file tools; R-33 Project configuration from a cloned repository; R-34 Tool calls auto-approved without full arguments. - R-29: A flawed or malicious server reaches SSH keys, cloud credentials, browser profiles and the home directory. - R-30: Arbitrary code runs with the client's privileges the moment the server starts. - R-31: Code runs on the machine with the user's privileges. - R-32: Server credentials and local files can be read and sent out through the tool. Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0). ## First controls to apply Start with the controls on the high rows, then work down. Controls name settings only where a primary source documents them; check each against the version your clients run. - R-29: Run local servers in a container or sandbox with minimal default privileges. - R-30: Show the exact command, with arguments and without truncation, before any server is added. - R-31: Allow only http and https schemes for authorization URLs and never open URLs through a shell. - R-32: Resolve paths on the server and reject any that leave an allowed directory. - R-33: Do not auto-enable project servers: leave enableAllProjectMcpServers unset or false and review each project server. - R-34: Require explicit confirmation for destructive, financial and data-sharing operations and show full parameters. - R-35: Keep the Inspector on a current version; releases 2.12.0, 2.14.0 and 2.15.0 carry security fixes. Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0). ## Execution and filesystem: 7 items ### R-29: Local server runs with the user's full privileges A local MCP server is an ordinary process launched by the host application. It inherits the user's environment variables and can read anything the account can read. **Impact:** A flawed or malicious server reaches SSH keys, cloud credentials, browser profiles and the home directory. **Controls:** - Run local servers in a container or sandbox with minimal default privileges. - Restrict filesystem access to the directories the server needs. - Run servers under a separate account or workspace where the client allows it. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### R-30: Malicious startup command in configuration A client configuration or one-click install carries a launch command that runs code. The spec lists a malicious startup command as a local-server compromise route. **Impact:** Arbitrary code runs with the client's privileges the moment the server starts. **Controls:** - Show the exact command, with arguments and without truncation, before any server is added. - Flag commands that use sudo, delete files, use network tools or reach SSH and home directories. - Distribute approved configurations through managed settings (VS Code allowedMcpServers, Claude Code managed-mcp.json). Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16). ### R-31: Command injection through arguments or URL opening Text from a model, a tool argument or a server-supplied OAuth authorization URL reaches a shell. The Inspector's 2.12.0 notes say a remote server's authorization URL could previously run commands through the opener on Windows and WSL. **Impact:** Code runs on the machine with the user's privileges. **Controls:** - Allow only http and https schemes for authorization URLs and never open URLs through a shell. - Never pass raw shell commands or unsanitized paths from tool arguments. - Validate tool inputs against strict schemas and allow-lists. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-32: Path traversal and arbitrary file read in file tools A tool that reads or uploads files accepts a path without limiting it to an intended directory. 14 of the 25 mcp-atlassian advisories published on 22 Sep 2026 concern file read through attachment tools. **Impact:** Server credentials and local files can be read and sent out through the tool. **Controls:** - Resolve paths on the server and reject any that leave an allowed directory. - Run file tools in a sandbox that mounts only the needed folders. - Upgrade mcp-atlassian to 0.22.0 or later. Severity: high Sources: [GHSA-93xw-j965-9mx3: mcp-atlassian arbitrary file read](https://github.com/advisories/GHSA-93xw-j965-9mx3) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-33: Project configuration from a cloned repository A repository carries MCP configuration that a client could load automatically. Claude Code's docs say a cloned repository cannot approve its own servers and that approval settings committed to the project are ignored in an untrusted folder. **Impact:** Opening a repository could start a server chosen by its author. **Controls:** - Do not auto-enable project servers: leave enableAllProjectMcpServers unset or false and review each project server. - Use workspace trust and require approval for project-scoped servers. - Lock server sets with managed-mcp.json and allowedMcpServers or deniedMcpServers. Severity: high Sources: [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16); [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings) (read 2026-11-16). ### R-34: Tool calls auto-approved without full arguments A client approves tool calls automatically, or shows only a tool name, so the user never sees what is about to run. **Impact:** A poisoned description or injected content can trigger destructive or data-sharing actions without anyone noticing. **Controls:** - Require explicit confirmation for destructive, financial and data-sharing operations and show full parameters. - Never auto-approve tool calls in multi-server setups. - Make sure the confirmation interface cannot be satisfied by text the model generates. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### R-35: Developer tools that spawn processes on a port A developer tool such as an MCP inspector can start stdio servers on request, and its API token is what protects that ability from other callers. **Impact:** A weak guard turns a testing tool into a way to run commands. The Inspector's own docs say its web backend can spawn processes and that stdio servers run as the user. **Controls:** - Keep the Inspector on a current version; releases 2.12.0, 2.14.0 and 2.15.0 carry security fixes. - Do not disable its authentication (DANGEROUSLY_OMIT_AUTH) outside an isolated test setup. - Bind it to loopback and never expose its port. Severity: medium Sources: [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16). ## Related pages - [The full MCP Risk Register](/mcp-risk-register/) - [Tool descriptions and schemas](/mcp-risk-register/tool-descriptions-and-schemas/) - [Data exposure and logging](/mcp-risk-register/data-exposure-and-logging/) - [Method](/mcp-risk-register/methodology/) - [MCP Security Checklist](/mcp-security-checklist/) --- # MCP data exposure and logging risks URL: https://mcpservercybersecurity.com/mcp-risk-register/data-exposure-and-logging/ Updated: 2026-11-16 Seven risks cover where secrets and data end up: inherited environment variables, plaintext token files, exfiltration through legitimate calls, secrets requested by elicitation, missing tool-call records, secrets inside logs and protected data returned through resource templates. Each row has controls and a primary source, read on 16 Nov 2026. ## What this dimension covers Where secrets and data end up: environment, token storage, exfiltration channels and logs. The register holds 7 rows here (R-36, R-37, R-38, R-39, R-40, R-41, R-42). Each row states what the risk is, its impact, concrete controls and at least one primary source. Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [GHSA-g5xv-mhgm-v5f6: mcp-atlassian plaintext OAuth token storage](https://github.com/advisories/GHSA-g5xv-mhgm-v5f6). ## Highest-severity rows Rows that meet the high rule under the written rule are: R-36 Secrets inherited through the environment; R-38 Data exfiltration through legitimate tool calls. - R-36: A server that needs one key can read all of them. - R-38: Data leaves through channels that policy allows. Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [GHSA-g5xv-mhgm-v5f6: mcp-atlassian plaintext OAuth token storage](https://github.com/advisories/GHSA-g5xv-mhgm-v5f6); [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security). ## First controls to apply Start with the controls on the high rows, then work down. Controls name settings only where a primary source documents them; check each against the version your clients run. - R-36: Start servers with a minimal environment and pass only the variables each one needs. - R-37: Use the OS credential store (macOS Keychain, Windows Credential Manager, Linux Secret Service). - R-38: Monitor for credentials or sensitive values from one server appearing in calls to another. - R-39: Reject servers that request secrets in form mode and use URL mode for credential steps. - R-40: Log every tool invocation with full parameters, user context and timestamp. - R-41: Redact secrets and personal data before storing tool-call logs. - R-42: Check authorization for the final resource that is returned, not the template. Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [GHSA-g5xv-mhgm-v5f6: mcp-atlassian plaintext OAuth token storage](https://github.com/advisories/GHSA-g5xv-mhgm-v5f6); [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security). ## Data exposure and logging: 7 items ### R-36: Secrets inherited through the environment A server started by the client inherits the environment of the host process, which often holds cloud keys and tokens. **Impact:** A server that needs one key can read all of them. **Controls:** - Start servers with a minimal environment and pass only the variables each one needs. - Use separate credentials per server with the least privilege. - Keep secrets in an OS credential store rather than in the shell profile. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-37: Tokens stored in plaintext A server or client saves OAuth tokens in a file with loose permissions or in plain text. mcp-atlassian advisories GHSA-g5xv-mhgm-v5f6 and GHSA-4596-2p6p-28cv describe this, and the Inspector's docs say its secrets fall back to a plaintext file when no OS keychain exists. **Impact:** Any local process or user that can read the file gets the tokens. **Controls:** - Use the OS credential store (macOS Keychain, Windows Credential Manager, Linux Secret Service). - Set file permissions to the owner only where files are unavoidable. - Never store tokens in MCP config files, which the OWASP cheat sheet also says. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [GHSA-g5xv-mhgm-v5f6: mcp-atlassian plaintext OAuth token storage](https://github.com/advisories/GHSA-g5xv-mhgm-v5f6) (read 2026-11-16); [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (read 2026-11-16). ### R-38: Data exfiltration through legitimate tool calls Injected instructions encode sensitive data into ordinary-looking calls, such as search queries or email subjects. **Impact:** Data leaves through channels that policy allows. **Controls:** - Monitor for credentials or sensitive values from one server appearing in calls to another. - Restrict which tools can send data outside the organization and require approval for them. - Limit the data each server can read in the first place. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-39: Sensitive data requested through form elicitation A server uses form-mode elicitation to ask the user for a password, API key or token, which passes through the client. **Impact:** Secrets pass through the client and its logs. The spec forbids it. **Controls:** - Reject servers that request secrets in form mode and use URL mode for credential steps. - Show which server is asking and let the user decline. - Review elicitation requests during server vetting. Severity: medium Sources: [MCP Elicitation, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/client/elicitation) (read 2026-11-16). ### R-40: No record of tool calls Tool invocations are not logged with their parameters, user and time, so there is nothing to investigate after an incident. **Impact:** A misuse or a poisoned call cannot be reconstructed, and anomalies go unseen. **Controls:** - Log every tool invocation with full parameters, user context and timestamp. - Send the logs to a SIEM and alert on new tools, admin-level queries and unusual call rates. - Keep the logs long enough to cover an investigation window you set. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### R-41: Secrets and personal data written into logs Logging full parameters also records secrets and personal data that appear in arguments and results. **Impact:** The log store becomes a second place to steal credentials from. **Controls:** - Redact secrets and personal data before storing tool-call logs. - Restrict who can read the logs and set a retention period. - Test the redaction rules with sample credentials. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### R-42: Protected data returned through resources and templates Resources and resource templates expose data by URI, and a template can retrieve a resource that a rule denied. **Impact:** Data outside a caller's rights is returned. FastMCP's GHSA-7xgj fix makes a denied concrete resource shadow a matching template. **Controls:** - Check authorization for the final resource that is returned, not the template. - Avoid overlapping proxy templates for protected resources. - Test reads for denied URIs through every route. Severity: medium Sources: [GHSA-7xgj-8392-5q4r: FastMCP post-transform authorization](https://github.com/advisories/GHSA-7xgj-8392-5q4r) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ## Related pages - [The full MCP Risk Register](/mcp-risk-register/) - [Execution and filesystem](/mcp-risk-register/execution-and-filesystem/) - [Supply chain](/mcp-risk-register/supply-chain/) - [Method](/mcp-risk-register/methodology/) - [MCP Security Checklist](/mcp-security-checklist/) --- # MCP Security Checklist URL: https://mcpservercybersecurity.com/mcp-security-checklist/ Updated: 2026-11-16 The MCP Security Checklist has 42 steps in six stages: discover, vet, authenticate, constrain, observe and respond. Each is a concrete action with why it matters, how to do it, a priority from one written rule (31 high, 11 medium, 0 low) and the risk identifiers it closes, with sources read on 16 Nov 2026. ## All 42 items **MCP Security Checklist: 42 items** | ID | Item | Category | Severity | |---|---|---|---| | [C-01](#C-01) | List every MCP client in use | [Discover](/mcp-security-checklist/discover/) | high | | [C-02](#C-02) | Collect MCP configuration files from machines | [Discover](/mcp-security-checklist/discover/) | high | | [C-03](#C-03) | Record how each server is launched | [Discover](/mcp-security-checklist/discover/) | high | | [C-04](#C-04) | Turn on endpoint discovery of configured servers | [Discover](/mcp-security-checklist/discover/) | high | | [C-05](#C-05) | Find remote and HTTP MCP endpoints on your network | [Discover](/mcp-security-checklist/discover/) | high | | [C-06](#C-06) | Assign an owner and review date to each server | [Discover](/mcp-security-checklist/discover/) | high | | [C-07](#C-07) | Compare the inventory with the approved catalog on a schedule | [Discover](/mcp-security-checklist/discover/) | high | | [C-08](#C-08) | Verify the publisher and the exact package name | [Vet](/mcp-security-checklist/vet/) | high | | [C-09](#C-09) | Diff the code against the upstream project | [Vet](/mcp-security-checklist/vet/) | high | | [C-10](#C-10) | Review every field of each tool definition | [Vet](/mcp-security-checklist/vet/) | high | | [C-11](#C-11) | Scan the server's dependencies | [Vet](/mcp-security-checklist/vet/) | medium | | [C-12](#C-12) | Check the SDK version against current advisories | [Vet](/mcp-security-checklist/vet/) | high | | [C-13](#C-13) | Test the server in a sandbox with network capture | [Vet](/mcp-security-checklist/vet/) | high | | [C-14](#C-14) | Review the launch manifest and environment variables | [Vet](/mcp-security-checklist/vet/) | high | | [C-15](#C-15) | Record the decision and pin the approved version | [Vet](/mcp-security-checklist/vet/) | medium | | [C-16](#C-16) | Require authentication on every HTTP MCP endpoint | [Authenticate](/mcp-security-checklist/authenticate/) | high | | [C-17](#C-17) | Validate the audience of every access token | [Authenticate](/mcp-security-checklist/authenticate/) | high | | [C-18](#C-18) | Never forward the received token downstream | [Authenticate](/mcp-security-checklist/authenticate/) | high | | [C-19](#C-19) | Use narrow scopes and a credential per server | [Authenticate](/mcp-security-checklist/authenticate/) | high | | [C-20](#C-20) | Bind stored credentials to their issuer | [Authenticate](/mcp-security-checklist/authenticate/) | high | | [C-21](#C-21) | Require per-client consent on MCP proxy servers | [Authenticate](/mcp-security-checklist/authenticate/) | medium | | [C-22](#C-22) | Store tokens in the operating system credential store | [Authenticate](/mcp-security-checklist/authenticate/) | medium | | [C-23](#C-23) | Allow only approved servers through managed settings | [Constrain](/mcp-security-checklist/constrain/) | high | | [C-24](#C-24) | Keep project servers from approving themselves | [Constrain](/mcp-security-checklist/constrain/) | high | | [C-25](#C-25) | Run local servers in a sandbox | [Constrain](/mcp-security-checklist/constrain/) | high | | [C-26](#C-26) | Start servers with a minimal environment | [Constrain](/mcp-security-checklist/constrain/) | high | | [C-27](#C-27) | Bind local HTTP servers to loopback and check Origin and Host | [Constrain](/mcp-security-checklist/constrain/) | high | | [C-28](#C-28) | Restrict outbound network access from servers | [Constrain](/mcp-security-checklist/constrain/) | high | | [C-29](#C-29) | Require confirmation for sensitive tool calls | [Constrain](/mcp-security-checklist/constrain/) | high | | [C-30](#C-30) | Check URLs before a client opens them | [Constrain](/mcp-security-checklist/constrain/) | high | | [C-31](#C-31) | Log every tool call with its parameters | [Observe](/mcp-security-checklist/observe/) | medium | | [C-32](#C-32) | Redact secrets and personal data in logs | [Observe](/mcp-security-checklist/observe/) | medium | | [C-33](#C-33) | Alert when tool definitions change | [Observe](/mcp-security-checklist/observe/) | medium | | [C-34](#C-34) | Alert on new servers, destinations and cross-server data flows | [Observe](/mcp-security-checklist/observe/) | high | | [C-35](#C-35) | Send MCP logs to the SIEM | [Observe](/mcp-security-checklist/observe/) | medium | | [C-36](#C-36) | Track advisories for approved components | [Observe](/mcp-security-checklist/observe/) | medium | | [C-37](#C-37) | Write a procedure to remove a server everywhere | [Respond](/mcp-security-checklist/respond/) | high | | [C-38](#C-38) | Rotate credentials the server could read | [Respond](/mcp-security-checklist/respond/) | high | | [C-39](#C-39) | Preserve logs and a copy of the package | [Respond](/mcp-security-checklist/respond/) | medium | | [C-40](#C-40) | Find affected machines from the inventory | [Respond](/mcp-security-checklist/respond/) | high | | [C-41](#C-41) | Patch and apply the settings that upgrades do not turn on | [Respond](/mcp-security-checklist/respond/) | high | | [C-42](#C-42) | Update vetting rules after each incident | [Respond](/mcp-security-checklist/respond/) | medium | ## What the checklist covers The checklist turns the register's risks into actions. It starts with finding every MCP server on machines, then vetting, authentication, constraint, observation and response. Where a client documents a setting, the step names it: VS Code Copilot enterprise-managed settings use allowedMcpServers and deniedMcpServers, and Claude Code documents managed-mcp.json and enableAllProjectMcpServers. Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security). ## How the steps are organized There are 42 steps with identifiers C-01 to C-42. The stages and their step counts are listed below. Each step lists the register rows it addresses in its controls and in a risks field. - Discover: 7 steps - Vet: 8 steps - Authenticate: 7 steps - Constrain: 8 steps - Observe: 6 steps - Respond: 6 steps Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Using the checklist in a rollout Teams with no inventory start at C-01 to C-07 and add C-23 once the catalog exists. Teams that already allow-list servers can start at vetting and observation. The tool on the home page tracks progress in the browser only. Endpoint tools that discover and control MCP servers can support the discover, constrain and observe stages. Vendors document different coverage, and the ranking on this site compares them. Sources: [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/); [DefenseClaw 1.0.0 release notes, Cisco](https://github.com/cisco-ai-defense/defenseclaw/releases/tag/1.0.0). ## Item detail ### C-01: List every MCP client in use Write down each application that can load MCP servers: IDEs, desktop AI apps and command-line agents, with versions. **Impact:** A server cannot be found if the client that loads it is unknown. **Controls:** - Survey developers and read software inventory for IDEs, desktop AI apps and agent command-line tools. - Record where each client keeps its MCP configuration (user, project and managed levels). - Register risks: R-07. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16). ### C-02: Collect MCP configuration files from machines Gather the config files that list servers, including project files that arrive with repositories. **Impact:** Config files hold the launch commands that decide what code runs. **Controls:** - Read user-level and project-level files such as .mcp.json and client settings files, and the managed settings you control. - Store the collected result with machine, user and date. - Register risks: R-07, R-33. Severity: high Sources: [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### C-03: Record how each server is launched For every server note the command, package name, version, publisher and any environment variables it receives. **Impact:** The launch command is what the machine executes; a changed package name or flag is the first sign of a problem. **Controls:** - Capture the full command without truncation, including npx or other launcher flags. - Note which environment variables and credentials the server is given. - Register risks: R-01, R-03, R-07. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### C-04: Turn on endpoint discovery of configured servers Use an endpoint tool that reports MCP servers configured on managed machines, so the inventory does not depend on surveys. **Impact:** Manual surveys go stale; machine-reported inventory shows servers as they are added. **Controls:** - Microsoft documents local AI agent discovery, generally available on Windows and macOS in October 2026, that reports each agent's configured MCP servers in Defender for Endpoint. - Snyk says Evo ADS MCP Governance discovers MCP servers across the estate, and Cisco's DefenseClaw 1.0.0 notes list a per-item MCP inventory; compare coverage against your clients before choosing. - Register risks: R-07. Severity: high Sources: [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16); [DefenseClaw 1.0.0 release notes, Cisco](https://github.com/cisco-ai-defense/defenseclaw/releases/tag/1.0.0) (read 2026-11-16). ### C-05: Find remote and HTTP MCP endpoints on your network List MCP servers that listen on a port or are hosted for team use, and who can reach them. **Impact:** A server listening on all interfaces, or without authentication, serves its tools to the network. **Controls:** - Check for MCP servers bound to 0.0.0.0 on developer subnets and for hosted servers in cloud accounts. - Record each endpoint's authentication method. - Register risks: R-09, R-17. Severity: high Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-06: Assign an owner and review date to each server Give every server in the inventory a named owner, a purpose and a date by which it is re-reviewed. **Impact:** Servers without an owner are never patched or removed. **Controls:** - Keep owner, purpose, version and next review in the catalog record. - Remove servers whose owner cannot be found. - Register risks: R-07. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16). ### C-07: Compare the inventory with the approved catalog on a schedule Diff what is installed against what is approved and report the difference to the owners. **Impact:** New and unapproved servers appear continuously; drift left unchecked becomes the norm. **Controls:** - Run the comparison weekly and report new servers by machine and team. - Track servers that changed version since approval. - Register risks: R-07, R-05. Severity: high Sources: [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-08: Verify the publisher and the exact package name Check the name, publisher and repository against the vendor's own documentation before the first install. **Impact:** Look-alike and repackaged packages differ from the real one by a character or an address. **Controls:** - Compare the package name letter by letter with the vendor's documented install command. - Check the publisher account, repository link and release history. - Register risks: R-01, R-02. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16). ### C-09: Diff the code against the upstream project For a server that copies an open-source project, compare it with upstream for changed endpoints and authentication code. **Impact:** Knostic described a package that copied Amazon's code and replaced the authentication and API addresses. **Controls:** - Search the code for hardcoded hosts and for changes to token or API endpoints. - Reject packages whose README does not account for the differences. - Register risks: R-02, R-03. Severity: high Sources: [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-10: Review every field of each tool definition Read the descriptions, parameter names, types, defaults, output schemas and annotations of every tool. **Impact:** The model reads all of these, and instructions can sit in any of them. **Controls:** - Look for text that tells the agent to read credentials, call other tools or hide actions from the user. - Run a scanner built for tool definitions as a second check; the OWASP cheat sheet names mcp-scan. - Register risks: R-23, R-25, R-26. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token) (read 2026-11-16); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/) (read 2026-11-16). ### C-11: Scan the server's dependencies Run dependency scanning on the server and review known vulnerabilities and maintenance state. **Impact:** A vulnerable dependency runs with the server's permissions. **Controls:** - Scan lockfiles and container images; track advisories for the servers and SDKs you approve. - Prefer servers with a documented update path. - Register risks: R-04. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16). ### C-12: Check the SDK version against current advisories Find which MCP SDK the server uses and compare it with the advisories published in October 2026. **Impact:** SDK flaws affect every server built on the affected versions. **Controls:** - Compare with the TypeScript SDK advisories (sdk 1.32.0 and server 2.3.0 are the fixed versions named) and the Python SDK advisories (mcp 1.30.0 and 2.2.0). - Check the fix needs a setting: GHSA-w4fh-qvv9-3v23 says upgrading alone changes nothing until the audience check is turned on. - Register risks: R-08, R-11, R-15, R-21. Severity: high Sources: [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (read 2026-11-16); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (read 2026-11-16); [GHSA-jm63-cq47-j5jp: MCP TypeScript server SDK cross-client messages](https://github.com/advisories/GHSA-jm63-cq47-j5jp) (read 2026-11-16); [GHSA-46jr-4wx8-v453: MCP TypeScript SDK memory exhaustion](https://github.com/advisories/GHSA-46jr-4wx8-v453) (read 2026-11-16). ### C-13: Test the server in a sandbox with network capture Run the candidate server in a disposable environment and record its files touched and hosts contacted. **Impact:** Behavior on first run shows whether the server does only what it says. **Controls:** - Use a container or virtual machine with no real credentials and log outbound connections. - Use the MCP Inspector to list tools and review definitions; keep it on a current release. - Register risks: R-29, R-22, R-35. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (read 2026-11-16). ### C-14: Review the launch manifest and environment variables Read the server's manifest or server.json entry and every environment variable it asks for. **Impact:** Variables such as NODE_OPTIONS or a registry setting change how the runtime behaves. **Controls:** - Reject entries with unexplained environment variables or runtime-control variables. - Inspector 2.15.0 flags NODE_OPTIONS and npm_config_registry in its server.json preview; use it as a model for your own review. - Register risks: R-30, R-03. Severity: high Sources: [MCP Inspector 2.15.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.15.0) (read 2026-11-16); [MCP Inspector 2.14.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.14.0) (read 2026-11-16). ### C-15: Record the decision and pin the approved version Store the approval with reviewer, date, version and a hash of the tool definitions. **Impact:** Without a pin, an approval does not describe tomorrow's server. **Controls:** - Pin the package version and record the hash of the full tool definition. - Do not use registry listings as the approval; key entries by name and version. - Register risks: R-05, R-06. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Registry v1.8.8 release notes](https://github.com/modelcontextprotocol/registry/releases/tag/v1.8.8) (read 2026-11-16). ### C-16: Require authentication on every HTTP MCP endpoint Make each remote or HTTP-exposed server refuse requests without valid credentials. **Impact:** The protocol makes authorization optional; a server that skips it serves anyone who reaches it. **Controls:** - Test that a request with no token and one with an arbitrary token are both refused. - Place the server on loopback or a private network unless authentication is in place. - Register risks: R-09. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [GHSA-wrhw-j3f9-8vc6: mcp-atlassian authentication bypass (CVE-2026-77244)](https://github.com/advisories/GHSA-wrhw-j3f9-8vc6) (read 2026-11-16). ### C-17: Validate the audience of every access token Reject tokens that were not issued for this server, using the resource indicator the client sent. **Impact:** A token for another service must not work here. **Controls:** - Check the aud claim against your server's canonical URI (RFC 8707), as the authorization spec requires. - Turn on the SDK's audience setting (for example validate_token_resource in the Python SDK) after upgrading. - Register risks: R-08. Severity: high Sources: [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (read 2026-11-16); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (read 2026-11-16); [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (read 2026-11-16). ### C-18: Never forward the received token downstream Obtain a separate token for each upstream API instead of passing the client's token on. **Impact:** Passing tokens through breaks audit trails and lets controls be bypassed. **Controls:** - Review code paths that call upstream APIs for pass-through of the Authorization header. - Log the identity of the caller and the upstream identity separately. - Register risks: R-08, R-10. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### C-19: Use narrow scopes and a credential per server Request the minimum scopes per tool, and give each server its own credential. **Impact:** One shared or broad token turns one flaw into access to everything it covers. **Controls:** - Prefer short-lived tokens over long-lived personal access tokens. - Review the scopes a server requests during vetting and refuse broad ones. - Register risks: R-12, R-36. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-20: Bind stored credentials to their issuer Make clients refuse to send saved credentials to an authorization server other than the one that issued them. **Impact:** Credentials sent to a server-named authorization server can be collected by that server. **Controls:** - Pass expectedIssuer or issuer where your SDK provides it, and clear credentials saved without one. - Reject authorization server metadata whose issuer does not match the expected origin. - Register risks: R-11, R-14. Severity: high Sources: [GHSA-6qxp-vccf-f47h: MCP TypeScript SDK OAuth client credentials (CVE-2026-104850)](https://github.com/advisories/GHSA-6qxp-vccf-f47h) (read 2026-11-16); [GHSA-vqqp-x5qm-5qp4: FastMCP client OAuth issuer binding](https://github.com/advisories/GHSA-vqqp-x5qm-5qp4) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### C-21: Require per-client consent on MCP proxy servers If you run an MCP proxy to a third-party API, store consent per user and per client and match redirect URIs exactly. **Impact:** Without per-client consent a proxy can be used to obtain authorization codes without approval. **Controls:** - Validate the OAuth state parameter and set consent cookies with the __Host- prefix, Secure, HttpOnly and SameSite=Lax. - Prefer Client ID Metadata Documents; Dynamic Client Registration is deprecated in 2026-07-28. - Register risks: R-10, R-14. Severity: medium Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (read 2026-11-16). ### C-22: Store tokens in the operating system credential store Keep access and refresh tokens in Keychain, Credential Manager or Secret Service, not in config files or plaintext. **Impact:** Files with loose permissions give tokens to any local reader. **Controls:** - Check where each client and server writes tokens; the Inspector's docs say its fallback is a plaintext file when no keychain exists. - Set owner-only permissions on any file that must hold a secret. - Register risks: R-37. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (read 2026-11-16); [GHSA-g5xv-mhgm-v5f6: mcp-atlassian plaintext OAuth token storage](https://github.com/advisories/GHSA-g5xv-mhgm-v5f6) (read 2026-11-16). ### C-23: Allow only approved servers through managed settings Push an allow list of servers to clients so unapproved servers do not load. **Impact:** A catalog that clients do not enforce is only advice. **Controls:** - VS Code Copilot enterprise-managed settings use allowedMcpServers and deniedMcpServers; an empty allowedMcpServers list blocks all servers and deny rules win over the allow list. - Claude Code supports managed-mcp.json for a fixed server set and allowedMcpServers and deniedMcpServers for restriction. - Register risks: R-01, R-07, R-33. Severity: high Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings) (read 2026-11-16); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16). ### C-24: Keep project servers from approving themselves Do not enable all project-scoped servers automatically, and require approval for servers that arrive in a repository. **Impact:** A cloned repository can carry configuration chosen by its author. **Controls:** - Leave enableAllProjectMcpServers unset or false in Claude Code and review each project server; Claude Code ignores such approvals committed to the project in an untrusted folder. - Use the client's workspace trust prompts. - Register risks: R-33, R-30. Severity: high Sources: [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### C-25: Run local servers in a sandbox Run each local server in a container or sandbox with minimal filesystem and network access. **Impact:** A local server otherwise has everything the user account has. **Controls:** - Mount only the directories the server needs and drop network access unless required. - Note that a bare chroot is not a sandbox, per the OWASP cheat sheet. - Register risks: R-29, R-32. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-26: Start servers with a minimal environment Pass each server only the variables it needs instead of the whole shell environment. **Impact:** Servers inherit the environment of the process that starts them. **Controls:** - Build the environment explicitly in the launch config. - Use separate credentials per server. - Register risks: R-36. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### C-27: Bind local HTTP servers to loopback and check Origin and Host Listen on 127.0.0.1, validate Origin and answer an invalid one with 403, and allow-list Host values. **Impact:** These checks close the DNS rebinding route to local services. **Controls:** - Follow the Streamable HTTP spec's security requirements for Origin and binding. - Prefer stdio or a Unix socket with restricted permissions for local servers. - Register risks: R-16, R-17. Severity: high Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16). ### C-28: Restrict outbound network access from servers Allow only the destinations a server needs, and route server-side fetches through an egress proxy. **Impact:** Blocked egress stops credential and data exfiltration and SSRF reaching internal hosts. **Controls:** - Block private, loopback and link-local ranges for fetches driven by URLs from servers or models; use a maintained library. - Alert on connections to destinations outside the allow list. - Register risks: R-22, R-18. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### C-29: Require confirmation for sensitive tool calls Show full parameters and ask for approval for destructive, financial and data-sharing operations, and never auto-approve in multi-server setups. **Impact:** Approvals are the control that sits between injected text and an action. **Controls:** - Use your own policy per tool rather than the server's annotations, which clients must treat as untrusted. - Make sure model-generated text cannot satisfy the confirmation. - Register risks: R-34, R-26, R-27. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### C-30: Check URLs before a client opens them In clients you build, allow only http and https for authorization URLs and open them without a shell. **Impact:** Server-supplied URLs reach the operating system's opener. **Controls:** - Reject javascript:, data:, file: and other schemes using an allow list. - Do not use cmd.exe, sh or PowerShell to open a URL. - Register risks: R-31, R-18. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16). ### C-31: Log every tool call with its parameters Record the tool, parameters, user and time for each invocation. **Impact:** Without records an incident cannot be reconstructed. **Controls:** - Capture the same fields for every client. - Keep the logs long enough for your investigation window. - Register risks: R-40. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### C-32: Redact secrets and personal data in logs Mask credentials and personal data in parameters and results before they are stored. **Impact:** Logs otherwise become a second store of secrets. **Controls:** - Test redaction with sample credentials. - Restrict log access and set retention. - Register risks: R-41. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-33: Alert when tool definitions change Compare tool definitions with the pinned hash and alert on any difference, including annotation changes. **Impact:** A changed definition is the signal of a rug pull. **Controls:** - Re-prompt for consent on change. - Alert on listChanged notifications from servers not expected to change. - Register risks: R-05, R-26. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### C-34: Alert on new servers, destinations and cross-server data flows Watch for unapproved servers appearing, new outbound hosts and one server's values showing up in another's calls. **Impact:** These patterns mark shadowing, exfiltration and unknown installs. **Controls:** - Add detections for new tools called, unusual call rates and admin-level queries. - Feed alerts to the same queue as other security alerts. - Register risks: R-24, R-38, R-22, R-07. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16). ### C-35: Send MCP logs to the SIEM Forward tool-call logs to the system your responders already use. **Impact:** Detection only works if the logs reach the people who can act. **Controls:** - Use a common schema across clients. - Map fields to user and device identity. - Register risks: R-40. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-36: Track advisories for approved components Subscribe to GitHub advisories and release notes for every server and SDK in the catalog. **Impact:** Fixes such as the October 2026 SDK advisories only matter if someone reads them. **Controls:** - Map each advisory to catalog entries and owners. - Record the date the fix was deployed. - Register risks: R-04, R-13. Severity: medium Sources: [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### C-37: Write a procedure to remove a server everywhere Define how to take a server out of the managed allow list and remove it from configurations. **Impact:** Speed matters once a server is found malicious. **Controls:** - Remove it from the allowed list first, then from config files and caches. - Test the procedure on a harmless server. - Register risks: R-07. Severity: high Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings) (read 2026-11-16); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16). ### C-38: Rotate credentials the server could read List the credentials in the server's environment and files and rotate them. **Impact:** Knostic's SP-API finding involves credentials sent to a hardcoded address; rotation limits the damage. **Controls:** - Rotate tokens, keys and refresh tokens that the server held. - Revoke sessions at the authorization server. - Register risks: R-36, R-02. Severity: high Sources: [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### C-39: Preserve logs and a copy of the package Keep tool-call logs and the exact package version for analysis. **Impact:** Registries remove malicious packages, and evidence goes with them. **Controls:** - Save the package, its hash and the config that launched it. - Export logs for the period the server was active. - Register risks: R-40. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-40: Find affected machines from the inventory Use the inventory to list every machine and user running the server or the affected version. **Impact:** Scope decides how many credentials to rotate. **Controls:** - Query by package name and version. - Include project-scoped configs. - Register risks: R-07. Severity: high Sources: [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16); [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint) (read 2026-11-16). ### C-41: Patch and apply the settings that upgrades do not turn on Upgrade affected servers and SDKs, then enable settings the fix needs. **Impact:** Some advisories say upgrading alone changes nothing. **Controls:** - Check the advisory text for required settings such as the audience check or expectedIssuer. - Re-test after the change. - Register risks: R-08, R-11. Severity: high Sources: [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (read 2026-11-16); [GHSA-6qxp-vccf-f47h: MCP TypeScript SDK OAuth client credentials (CVE-2026-104850)](https://github.com/advisories/GHSA-6qxp-vccf-f47h) (read 2026-11-16). ### C-42: Update vetting rules after each incident Add what you learned to the vetting checks and catalog rules. **Impact:** Each incident shows a gap in review. **Controls:** - Record the cause and the control that would have caught it. - Update the checklist edition and note the date. - Register risks: R-06. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ## Download and cite [Download the MCP Security Checklist as CSV or JSON](/datasets/mcp-security-checklist/). Cite it as: Cite assets as: MCP Server Security, "", edition 16 Nov 2026, https://mcpservercybersecurity.com//, licensed CC BY 4.0. ## Related pages - [How the MCP Security Checklist is built](/mcp-security-checklist/methodology/) - [MCP Security Checklist data download](/datasets/mcp-security-checklist/) - [Discover](/mcp-security-checklist/discover/) - [Vet](/mcp-security-checklist/vet/) - [Authenticate](/mcp-security-checklist/authenticate/) - [Constrain](/mcp-security-checklist/constrain/) - [Observe](/mcp-security-checklist/observe/) - [Respond](/mcp-security-checklist/respond/) - [MCP Risk Register](/mcp-risk-register/) - [The ranking of tools](/ranking/) --- # How the MCP Security Checklist is built URL: https://mcpservercybersecurity.com/mcp-security-checklist/methodology/ Updated: 2026-11-16 The checklist has 42 steps in six stages, discover, vet, authenticate, constrain, observe and respond. Each step maps to risk identifiers in the MCP Risk Register and carries a priority set by one written rule, with sources read on 16 Nov 2026. It is released under CC BY 4.0 with citation and corrections guidance below. ## What the checklist is The checklist is 42 actions that a platform or security engineer can tick, ordered in six stages. Each step has what to do, why it matters, how to do it and the register risks it closes. Steps contain no exploit code or bypass instructions. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security). ## Sources and read date Steps are drawn from the MCP specification's security pages, the OWASP MCP Security Cheat Sheet, GitHub advisories, maintainer release notes and vendor documentation for managed settings. All pages were fetched on 16 Nov 2026 (read date 2026-11-16) and only pages that returned HTTP 200 are cited. Where a step names a setting, such as allowedMcpServers in VS Code Copilot enterprise-managed settings or managed-mcp.json in Claude Code, the vendor's own documentation is cited. Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Priority rule Each step carries the priority of the highest-severity register risk it addresses: high if it addresses any risk the register marks high, medium if the highest is medium, low if the highest is low. The severity rule for the register rows is stated on the register's method page. Start with high steps; the order inside a dimension is the order to do them. On 16 Nov 2026 the checklist has 31 high, 11 medium and 0 low steps. Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices). ## How to cite Cite as: MCP Security Checklist, MCP Server Security, edition of 16 Nov 2026, https://mcpservercybersecurity.com/mcp-security-checklist/, licensed CC BY 4.0. Cite a step by its identifier (for example C-17) and the edition date. ## License The checklist text and the downloadable CSV and JSON files are released under Creative Commons Attribution 4.0 International (CC BY 4.0). You may copy and adapt them with credit and a link to the license. ## Corrections If a setting name changed, a source moved or a step is wrong, use the corrections page on this site. Corrections are dated and listed there, and the edition date changes when steps change. ## Counts - 42 items in 6 categories. - 31 with severity high. - 11 with severity medium. ## Related pages - [The MCP Security Checklist](/mcp-security-checklist/) - [MCP Security Checklist download](/datasets/mcp-security-checklist/) - [Scoring method](/method/) - [Discover](/mcp-security-checklist/discover/) - [Vet](/mcp-security-checklist/vet/) - [Authenticate](/mcp-security-checklist/authenticate/) --- # MCP security checklist: discover URL: https://mcpservercybersecurity.com/mcp-security-checklist/discover/ Updated: 2026-11-16 Seven steps find every MCP client and server on machines before anything else: list clients, collect config files, record launch commands, turn on endpoint discovery, find network listeners, assign owners and diff against the catalog. Each step names controls and sources, with priority set by the written rule, read on 16 Nov 2026. ## What this stage covers Find every MCP client and server on machines and in repositories. There are 7 steps here (C-01 to C-07), each a concrete action an engineer can tick. Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices). ## Steps in order Do these in order; each step lists how to do it and the register risks it closes. - C-01 List every MCP client in use (priority high) - C-02 Collect MCP configuration files from machines (priority high) - C-03 Record how each server is launched (priority high) - C-04 Turn on endpoint discovery of configured servers (priority high) - C-05 Find remote and HTTP MCP endpoints on your network (priority high) - C-06 Assign an owner and review date to each server (priority high) - C-07 Compare the inventory with the approved catalog on a schedule (priority high) Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint). ## Where the settings come from Setting names appear only where a primary source documents them. Check each against the client version in use before rollout, because the 2026-07-28 MCP revision and client releases changed several defaults. Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint). ## Discover: 7 items ### C-01: List every MCP client in use Write down each application that can load MCP servers: IDEs, desktop AI apps and command-line agents, with versions. **Impact:** A server cannot be found if the client that loads it is unknown. **Controls:** - Survey developers and read software inventory for IDEs, desktop AI apps and agent command-line tools. - Record where each client keeps its MCP configuration (user, project and managed levels). - Register risks: R-07. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16). ### C-02: Collect MCP configuration files from machines Gather the config files that list servers, including project files that arrive with repositories. **Impact:** Config files hold the launch commands that decide what code runs. **Controls:** - Read user-level and project-level files such as .mcp.json and client settings files, and the managed settings you control. - Store the collected result with machine, user and date. - Register risks: R-07, R-33. Severity: high Sources: [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### C-03: Record how each server is launched For every server note the command, package name, version, publisher and any environment variables it receives. **Impact:** The launch command is what the machine executes; a changed package name or flag is the first sign of a problem. **Controls:** - Capture the full command without truncation, including npx or other launcher flags. - Note which environment variables and credentials the server is given. - Register risks: R-01, R-03, R-07. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### C-04: Turn on endpoint discovery of configured servers Use an endpoint tool that reports MCP servers configured on managed machines, so the inventory does not depend on surveys. **Impact:** Manual surveys go stale; machine-reported inventory shows servers as they are added. **Controls:** - Microsoft documents local AI agent discovery, generally available on Windows and macOS in October 2026, that reports each agent's configured MCP servers in Defender for Endpoint. - Snyk says Evo ADS MCP Governance discovers MCP servers across the estate, and Cisco's DefenseClaw 1.0.0 notes list a per-item MCP inventory; compare coverage against your clients before choosing. - Register risks: R-07. Severity: high Sources: [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16); [DefenseClaw 1.0.0 release notes, Cisco](https://github.com/cisco-ai-defense/defenseclaw/releases/tag/1.0.0) (read 2026-11-16). ### C-05: Find remote and HTTP MCP endpoints on your network List MCP servers that listen on a port or are hosted for team use, and who can reach them. **Impact:** A server listening on all interfaces, or without authentication, serves its tools to the network. **Controls:** - Check for MCP servers bound to 0.0.0.0 on developer subnets and for hosted servers in cloud accounts. - Record each endpoint's authentication method. - Register risks: R-09, R-17. Severity: high Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-06: Assign an owner and review date to each server Give every server in the inventory a named owner, a purpose and a date by which it is re-reviewed. **Impact:** Servers without an owner are never patched or removed. **Controls:** - Keep owner, purpose, version and next review in the catalog record. - Remove servers whose owner cannot be found. - Register risks: R-07. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16). ### C-07: Compare the inventory with the approved catalog on a schedule Diff what is installed against what is approved and report the difference to the owners. **Impact:** New and unapproved servers appear continuously; drift left unchecked becomes the norm. **Controls:** - Run the comparison weekly and report new servers by machine and team. - Track servers that changed version since approval. - Register risks: R-07, R-05. Severity: high Sources: [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ## Related pages - [The full MCP Security Checklist](/mcp-security-checklist/) - [Respond](/mcp-security-checklist/respond/) - [Vet](/mcp-security-checklist/vet/) - [Method](/mcp-security-checklist/methodology/) - [MCP Risk Register](/mcp-risk-register/) --- # MCP security checklist: vet URL: https://mcpservercybersecurity.com/mcp-security-checklist/vet/ Updated: 2026-11-16 Eight steps decide whether a server may run: verify the publisher and package name, diff against upstream, review every tool field, scan dependencies, check the SDK against advisories, test in a sandbox, read the launch manifest and record a pinned approval. Each carries controls and sources, read on 16 Nov 2026. ## What this stage covers Decide whether a server is fit to run before it runs. There are 8 steps here (C-08 to C-15), each a concrete action an engineer can tick. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token). ## Steps in order Do these in order; each step lists how to do it and the register risks it closes. - C-08 Verify the publisher and the exact package name (priority high) - C-09 Diff the code against the upstream project (priority high) - C-10 Review every field of each tool definition (priority high) - C-11 Scan the server's dependencies (priority medium) - C-12 Check the SDK version against current advisories (priority high) - C-13 Test the server in a sandbox with network capture (priority high) - C-14 Review the launch manifest and environment variables (priority high) - C-15 Record the decision and pin the approved version (priority medium) Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/). ## Where the settings come from Setting names appear only where a primary source documents them. Check each against the client version in use before rollout, because the 2026-07-28 MCP revision and client releases changed several defaults. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/). ## Vet: 8 items ### C-08: Verify the publisher and the exact package name Check the name, publisher and repository against the vendor's own documentation before the first install. **Impact:** Look-alike and repackaged packages differ from the real one by a character or an address. **Controls:** - Compare the package name letter by letter with the vendor's documented install command. - Check the publisher account, repository link and release history. - Register risks: R-01, R-02. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16). ### C-09: Diff the code against the upstream project For a server that copies an open-source project, compare it with upstream for changed endpoints and authentication code. **Impact:** Knostic described a package that copied Amazon's code and replaced the authentication and API addresses. **Controls:** - Search the code for hardcoded hosts and for changes to token or API endpoints. - Reject packages whose README does not account for the differences. - Register risks: R-02, R-03. Severity: high Sources: [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-10: Review every field of each tool definition Read the descriptions, parameter names, types, defaults, output schemas and annotations of every tool. **Impact:** The model reads all of these, and instructions can sit in any of them. **Controls:** - Look for text that tells the agent to read credentials, call other tools or hide actions from the user. - Run a scanner built for tool definitions as a second check; the OWASP cheat sheet names mcp-scan. - Register risks: R-23, R-25, R-26. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token) (read 2026-11-16); [AI tool poisoning through MCP tool descriptions, CrowdStrike](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/) (read 2026-11-16). ### C-11: Scan the server's dependencies Run dependency scanning on the server and review known vulnerabilities and maintenance state. **Impact:** A vulnerable dependency runs with the server's permissions. **Controls:** - Scan lockfiles and container images; track advisories for the servers and SDKs you approve. - Prefer servers with a documented update path. - Register risks: R-04. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16). ### C-12: Check the SDK version against current advisories Find which MCP SDK the server uses and compare it with the advisories published in October 2026. **Impact:** SDK flaws affect every server built on the affected versions. **Controls:** - Compare with the TypeScript SDK advisories (sdk 1.32.0 and server 2.3.0 are the fixed versions named) and the Python SDK advisories (mcp 1.30.0 and 2.2.0). - Check the fix needs a setting: GHSA-w4fh-qvv9-3v23 says upgrading alone changes nothing until the audience check is turned on. - Register risks: R-08, R-11, R-15, R-21. Severity: high Sources: [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (read 2026-11-16); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (read 2026-11-16); [GHSA-jm63-cq47-j5jp: MCP TypeScript server SDK cross-client messages](https://github.com/advisories/GHSA-jm63-cq47-j5jp) (read 2026-11-16); [GHSA-46jr-4wx8-v453: MCP TypeScript SDK memory exhaustion](https://github.com/advisories/GHSA-46jr-4wx8-v453) (read 2026-11-16). ### C-13: Test the server in a sandbox with network capture Run the candidate server in a disposable environment and record its files touched and hosts contacted. **Impact:** Behavior on first run shows whether the server does only what it says. **Controls:** - Use a container or virtual machine with no real credentials and log outbound connections. - Use the MCP Inspector to list tools and review definitions; keep it on a current release. - Register risks: R-29, R-22, R-35. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (read 2026-11-16). ### C-14: Review the launch manifest and environment variables Read the server's manifest or server.json entry and every environment variable it asks for. **Impact:** Variables such as NODE_OPTIONS or a registry setting change how the runtime behaves. **Controls:** - Reject entries with unexplained environment variables or runtime-control variables. - Inspector 2.15.0 flags NODE_OPTIONS and npm_config_registry in its server.json preview; use it as a model for your own review. - Register risks: R-30, R-03. Severity: high Sources: [MCP Inspector 2.15.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.15.0) (read 2026-11-16); [MCP Inspector 2.14.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.14.0) (read 2026-11-16). ### C-15: Record the decision and pin the approved version Store the approval with reviewer, date, version and a hash of the tool definitions. **Impact:** Without a pin, an approval does not describe tomorrow's server. **Controls:** - Pin the package version and record the hash of the full tool definition. - Do not use registry listings as the approval; key entries by name and version. - Register risks: R-05, R-06. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Registry v1.8.8 release notes](https://github.com/modelcontextprotocol/registry/releases/tag/v1.8.8) (read 2026-11-16). ## Related pages - [The full MCP Security Checklist](/mcp-security-checklist/) - [Discover](/mcp-security-checklist/discover/) - [Authenticate](/mcp-security-checklist/authenticate/) - [Method](/mcp-security-checklist/methodology/) - [MCP Risk Register](/mcp-risk-register/) --- # MCP security checklist: authenticate URL: https://mcpservercybersecurity.com/mcp-security-checklist/authenticate/ Updated: 2026-11-16 Seven steps make sure callers and tokens are the right ones: require authentication on HTTP endpoints, validate token audience, avoid token passthrough, use narrow scopes, bind credentials to issuers, require proxy consent and store tokens in the operating system credential store. Each cites the MCP spec or an advisory, read on 16 Nov 2026. ## What this stage covers Make sure callers and tokens are the right ones. There are 7 steps here (C-16 to C-22), each a concrete action an engineer can tick. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [GHSA-wrhw-j3f9-8vc6: mcp-atlassian authentication bypass (CVE-2026-77244)](https://github.com/advisories/GHSA-wrhw-j3f9-8vc6); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization). ## Steps in order Do these in order; each step lists how to do it and the register risks it closes. - C-16 Require authentication on every HTTP MCP endpoint (priority high) - C-17 Validate the audience of every access token (priority high) - C-18 Never forward the received token downstream (priority high) - C-19 Use narrow scopes and a credential per server (priority high) - C-20 Bind stored credentials to their issuer (priority high) - C-21 Require per-client consent on MCP proxy servers (priority medium) - C-22 Store tokens in the operating system credential store (priority medium) Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [GHSA-wrhw-j3f9-8vc6: mcp-atlassian authentication bypass (CVE-2026-77244)](https://github.com/advisories/GHSA-wrhw-j3f9-8vc6); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23). ## Where the settings come from Setting names appear only where a primary source documents them. Check each against the client version in use before rollout, because the 2026-07-28 MCP revision and client releases changed several defaults. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [GHSA-wrhw-j3f9-8vc6: mcp-atlassian authentication bypass (CVE-2026-77244)](https://github.com/advisories/GHSA-wrhw-j3f9-8vc6); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23). ## Authenticate: 7 items ### C-16: Require authentication on every HTTP MCP endpoint Make each remote or HTTP-exposed server refuse requests without valid credentials. **Impact:** The protocol makes authorization optional; a server that skips it serves anyone who reaches it. **Controls:** - Test that a request with no token and one with an arbitrary token are both refused. - Place the server on loopback or a private network unless authentication is in place. - Register risks: R-09. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [GHSA-wrhw-j3f9-8vc6: mcp-atlassian authentication bypass (CVE-2026-77244)](https://github.com/advisories/GHSA-wrhw-j3f9-8vc6) (read 2026-11-16). ### C-17: Validate the audience of every access token Reject tokens that were not issued for this server, using the resource indicator the client sent. **Impact:** A token for another service must not work here. **Controls:** - Check the aud claim against your server's canonical URI (RFC 8707), as the authorization spec requires. - Turn on the SDK's audience setting (for example validate_token_resource in the Python SDK) after upgrading. - Register risks: R-08. Severity: high Sources: [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (read 2026-11-16); [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (read 2026-11-16); [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (read 2026-11-16). ### C-18: Never forward the received token downstream Obtain a separate token for each upstream API instead of passing the client's token on. **Impact:** Passing tokens through breaks audit trails and lets controls be bypassed. **Controls:** - Review code paths that call upstream APIs for pass-through of the Authorization header. - Log the identity of the caller and the upstream identity separately. - Register risks: R-08, R-10. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### C-19: Use narrow scopes and a credential per server Request the minimum scopes per tool, and give each server its own credential. **Impact:** One shared or broad token turns one flaw into access to everything it covers. **Controls:** - Prefer short-lived tokens over long-lived personal access tokens. - Review the scopes a server requests during vetting and refuse broad ones. - Register risks: R-12, R-36. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-20: Bind stored credentials to their issuer Make clients refuse to send saved credentials to an authorization server other than the one that issued them. **Impact:** Credentials sent to a server-named authorization server can be collected by that server. **Controls:** - Pass expectedIssuer or issuer where your SDK provides it, and clear credentials saved without one. - Reject authorization server metadata whose issuer does not match the expected origin. - Register risks: R-11, R-14. Severity: high Sources: [GHSA-6qxp-vccf-f47h: MCP TypeScript SDK OAuth client credentials (CVE-2026-104850)](https://github.com/advisories/GHSA-6qxp-vccf-f47h) (read 2026-11-16); [GHSA-vqqp-x5qm-5qp4: FastMCP client OAuth issuer binding](https://github.com/advisories/GHSA-vqqp-x5qm-5qp4) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ### C-21: Require per-client consent on MCP proxy servers If you run an MCP proxy to a third-party API, store consent per user and per client and match redirect URIs exactly. **Impact:** Without per-client consent a proxy can be used to obtain authorization codes without approval. **Controls:** - Validate the OAuth state parameter and set consent cookies with the __Host- prefix, Secure, HttpOnly and SameSite=Lax. - Prefer Client ID Metadata Documents; Dynamic Client Registration is deprecated in 2026-07-28. - Register risks: R-10, R-14. Severity: medium Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Authorization, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (read 2026-11-16). ### C-22: Store tokens in the operating system credential store Keep access and refresh tokens in Keychain, Credential Manager or Secret Service, not in config files or plaintext. **Impact:** Files with loose permissions give tokens to any local reader. **Controls:** - Check where each client and server writes tokens; the Inspector's docs say its fallback is a plaintext file when no keychain exists. - Set owner-only permissions on any file that must hold a secret. - Register risks: R-37. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (read 2026-11-16); [GHSA-g5xv-mhgm-v5f6: mcp-atlassian plaintext OAuth token storage](https://github.com/advisories/GHSA-g5xv-mhgm-v5f6) (read 2026-11-16). ## Related pages - [The full MCP Security Checklist](/mcp-security-checklist/) - [Vet](/mcp-security-checklist/vet/) - [Constrain](/mcp-security-checklist/constrain/) - [Method](/mcp-security-checklist/methodology/) - [MCP Risk Register](/mcp-risk-register/) --- # MCP security checklist: constrain URL: https://mcpservercybersecurity.com/mcp-security-checklist/constrain/ Updated: 2026-11-16 Eight steps limit what approved servers can do: managed allow lists, project approval settings, sandboxes, minimal environments, loopback binding with Origin checks, egress limits, confirmation for sensitive calls and safe URL opening. Settings named, such as allowedMcpServers and managed-mcp.json, come from vendor docs read on 16 Nov 2026. ## What this stage covers Limit what approved servers can run, reach and send. There are 8 steps here (C-23 to C-30), each a concrete action an engineer can tick. Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices). ## Steps in order Do these in order; each step lists how to do it and the register risks it closes. - C-23 Allow only approved servers through managed settings (priority high) - C-24 Keep project servers from approving themselves (priority high) - C-25 Run local servers in a sandbox (priority high) - C-26 Start servers with a minimal environment (priority high) - C-27 Bind local HTTP servers to loopback and check Origin and Host (priority high) - C-28 Restrict outbound network access from servers (priority high) - C-29 Require confirmation for sensitive tool calls (priority high) - C-30 Check URLs before a client opens them (priority high) Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security). ## Where the settings come from Setting names appear only where a primary source documents them. Check each against the client version in use before rollout, because the 2026-07-28 MCP revision and client releases changed several defaults. Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security). ## Constrain: 8 items ### C-23: Allow only approved servers through managed settings Push an allow list of servers to clients so unapproved servers do not load. **Impact:** A catalog that clients do not enforce is only advice. **Controls:** - VS Code Copilot enterprise-managed settings use allowedMcpServers and deniedMcpServers; an empty allowedMcpServers list blocks all servers and deny rules win over the allow list. - Claude Code supports managed-mcp.json for a fixed server set and allowedMcpServers and deniedMcpServers for restriction. - Register risks: R-01, R-07, R-33. Severity: high Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings) (read 2026-11-16); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16). ### C-24: Keep project servers from approving themselves Do not enable all project-scoped servers automatically, and require approval for servers that arrive in a repository. **Impact:** A cloned repository can carry configuration chosen by its author. **Controls:** - Leave enableAllProjectMcpServers unset or false in Claude Code and review each project server; Claude Code ignores such approvals committed to the project in an untrusted folder. - Use the client's workspace trust prompts. - Register risks: R-33, R-30. Severity: high Sources: [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### C-25: Run local servers in a sandbox Run each local server in a container or sandbox with minimal filesystem and network access. **Impact:** A local server otherwise has everything the user account has. **Controls:** - Mount only the directories the server needs and drop network access unless required. - Note that a bare chroot is not a sandbox, per the OWASP cheat sheet. - Register risks: R-29, R-32. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-26: Start servers with a minimal environment Pass each server only the variables it needs instead of the whole shell environment. **Impact:** Servers inherit the environment of the process that starts them. **Controls:** - Build the environment explicitly in the launch config. - Use separate credentials per server. - Register risks: R-36. Severity: high Sources: [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### C-27: Bind local HTTP servers to loopback and check Origin and Host Listen on 127.0.0.1, validate Origin and answer an invalid one with 403, and allow-list Host values. **Impact:** These checks close the DNS rebinding route to local services. **Controls:** - Follow the Streamable HTTP spec's security requirements for Origin and binding. - Prefer stdio or a Unix socket with restricted permissions for local servers. - Register risks: R-16, R-17. Severity: high Sources: [MCP Streamable HTTP transport, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16). ### C-28: Restrict outbound network access from servers Allow only the destinations a server needs, and route server-side fetches through an egress proxy. **Impact:** Blocked egress stops credential and data exfiltration and SSRF reaching internal hosts. **Controls:** - Block private, loopback and link-local ranges for fetches driven by URLs from servers or models; use a maintained library. - Alert on connections to destinations outside the allow list. - Register risks: R-22, R-18. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### C-29: Require confirmation for sensitive tool calls Show full parameters and ask for approval for destructive, financial and data-sharing operations, and never auto-approve in multi-server setups. **Impact:** Approvals are the control that sits between injected text and an action. **Controls:** - Use your own policy per tool rather than the server's annotations, which clients must treat as untrusted. - Make sure model-generated text cannot satisfy the confirmation. - Register risks: R-34, R-26, R-27. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### C-30: Check URLs before a client opens them In clients you build, allow only http and https for authorization URLs and open them without a shell. **Impact:** Server-supplied URLs reach the operating system's opener. **Controls:** - Reject javascript:, data:, file: and other schemes using an allow list. - Do not use cmd.exe, sh or PowerShell to open a URL. - Register risks: R-31, R-18. Severity: high Sources: [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16); [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (read 2026-11-16). ## Related pages - [The full MCP Security Checklist](/mcp-security-checklist/) - [Authenticate](/mcp-security-checklist/authenticate/) - [Observe](/mcp-security-checklist/observe/) - [Method](/mcp-security-checklist/methodology/) - [MCP Risk Register](/mcp-risk-register/) --- # MCP security checklist: observe URL: https://mcpservercybersecurity.com/mcp-security-checklist/observe/ Updated: 2026-11-16 Six steps keep a record and watch for change: log every tool call with parameters, redact secrets, alert on changed tool definitions, alert on new servers and destinations, forward logs to the SIEM and track advisories for approved components. Each step has controls and sources and a priority from the written rule, read on 16 Nov 2026. ## What this stage covers Record and watch tool calls and changes. There are 6 steps here (C-31 to C-36), each a concrete action an engineer can tick. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools). ## Steps in order Do these in order; each step lists how to do it and the register risks it closes. - C-31 Log every tool call with its parameters (priority medium) - C-32 Redact secrets and personal data in logs (priority medium) - C-33 Alert when tool definitions change (priority medium) - C-34 Alert on new servers, destinations and cross-server data flows (priority high) - C-35 Send MCP logs to the SIEM (priority medium) - C-36 Track advisories for approved components (priority medium) Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/). ## Where the settings come from Setting names appear only where a primary source documents them. Check each against the client version in use before rollout, because the 2026-07-28 MCP revision and client releases changed several defaults. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/). ## Observe: 6 items ### C-31: Log every tool call with its parameters Record the tool, parameters, user and time for each invocation. **Impact:** Without records an incident cannot be reconstructed. **Controls:** - Capture the same fields for every client. - Keep the logs long enough for your investigation window. - Register risks: R-40. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (read 2026-11-16). ### C-32: Redact secrets and personal data in logs Mask credentials and personal data in parameters and results before they are stored. **Impact:** Logs otherwise become a second store of secrets. **Controls:** - Test redaction with sample credentials. - Restrict log access and set retention. - Register risks: R-41. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-33: Alert when tool definitions change Compare tool definitions with the pinned hash and alert on any difference, including annotation changes. **Impact:** A changed definition is the signal of a rug pull. **Controls:** - Re-prompt for consent on change. - Alert on listChanged notifications from servers not expected to change. - Register risks: R-05, R-26. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [MCP Tools, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (read 2026-11-16). ### C-34: Alert on new servers, destinations and cross-server data flows Watch for unapproved servers appearing, new outbound hosts and one server's values showing up in another's calls. **Impact:** These patterns mark shadowing, exfiltration and unknown installs. **Controls:** - Add detections for new tools called, unusual call rates and admin-level queries. - Feed alerts to the same queue as other security alerts. - Register risks: R-24, R-38, R-22, R-07. Severity: high Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16); [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16). ### C-35: Send MCP logs to the SIEM Forward tool-call logs to the system your responders already use. **Impact:** Detection only works if the logs reach the people who can act. **Controls:** - Use a common schema across clients. - Map fields to user and device identity. - Register risks: R-40. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-36: Track advisories for approved components Subscribe to GitHub advisories and release notes for every server and SDK in the catalog. **Impact:** Fixes such as the October 2026 SDK advisories only matter if someone reads them. **Controls:** - Map each advisory to catalog entries and owners. - Record the date the fix was deployed. - Register risks: R-04, R-13. Severity: medium Sources: [GHSA-rvq5-wwqv-78pq: MCP TypeScript SDK bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (read 2026-11-16); [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (read 2026-11-16). ## Related pages - [The full MCP Security Checklist](/mcp-security-checklist/) - [Constrain](/mcp-security-checklist/constrain/) - [Respond](/mcp-security-checklist/respond/) - [Method](/mcp-security-checklist/methodology/) - [MCP Risk Register](/mcp-risk-register/) --- # MCP security checklist: respond URL: https://mcpservercybersecurity.com/mcp-security-checklist/respond/ Updated: 2026-11-16 Six steps cover a bad or vulnerable server: remove it everywhere, rotate credentials it could read, preserve logs and the package, find affected machines from the inventory, patch and apply settings an upgrade does not turn on, and update vetting rules. Each cites sources, with priority from the written rule, read on 16 Nov 2026. ## What this stage covers Act when a server is found bad or vulnerable. There are 6 steps here (C-37 to C-42), each a concrete action an engineer can tick. Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip). ## Steps in order Do these in order; each step lists how to do it and the register risks it closes. - C-37 Write a procedure to remove a server everywhere (priority high) - C-38 Rotate credentials the server could read (priority high) - C-39 Preserve logs and a copy of the package (priority medium) - C-40 Find affected machines from the inventory (priority high) - C-41 Patch and apply the settings that upgrades do not turn on (priority high) - C-42 Update vetting rules after each incident (priority medium) Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security). ## Where the settings come from Setting names appear only where a primary source documents them. Check each against the client version in use before rollout, because the 2026-07-28 MCP revision and client releases changed several defaults. Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp); [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security). ## Respond: 6 items ### C-37: Write a procedure to remove a server everywhere Define how to take a server out of the managed allow list and remove it from configurations. **Impact:** Speed matters once a server is found malicious. **Controls:** - Remove it from the allowed list first, then from config files and caches. - Test the procedure on a harmless server. - Register risks: R-07. Severity: high Sources: [AI settings for enterprise, Visual Studio Code docs](https://code.visualstudio.com/docs/enterprise/ai-settings) (read 2026-11-16); [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (read 2026-11-16). ### C-38: Rotate credentials the server could read List the credentials in the server's environment and files and rotate them. **Impact:** Knostic's SP-API finding involves credentials sent to a hardcoded address; rotation limits the damage. **Controls:** - Rotate tokens, keys and refresh tokens that the server held. - Revoke sessions at the authorization server. - Register risks: R-36, R-02. Severity: high Sources: [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (read 2026-11-16); [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (read 2026-11-16). ### C-39: Preserve logs and a copy of the package Keep tool-call logs and the exact package version for analysis. **Impact:** Registries remove malicious packages, and evidence goes with them. **Controls:** - Save the package, its hash and the config that launched it. - Export logs for the period the server was active. - Register risks: R-40. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ### C-40: Find affected machines from the inventory Use the inventory to list every machine and user running the server or the affected version. **Impact:** Scope decides how many credentials to rotate. **Controls:** - Query by package name and version. - Include project-scoped configs. - Register risks: R-07. Severity: high Sources: [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (read 2026-11-16); [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint) (read 2026-11-16). ### C-41: Patch and apply the settings that upgrades do not turn on Upgrade affected servers and SDKs, then enable settings the fix needs. **Impact:** Some advisories say upgrading alone changes nothing. **Controls:** - Check the advisory text for required settings such as the audience check or expectedIssuer. - Re-test after the change. - Register risks: R-08, R-11. Severity: high Sources: [GHSA-w4fh-qvv9-3v23: MCP Python SDK bearer-token audience](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (read 2026-11-16); [GHSA-6qxp-vccf-f47h: MCP TypeScript SDK OAuth client credentials (CVE-2026-104850)](https://github.com/advisories/GHSA-6qxp-vccf-f47h) (read 2026-11-16). ### C-42: Update vetting rules after each incident Add what you learned to the vetting checks and catalog rules. **Impact:** Each incident shows a gap in review. **Controls:** - Record the cause and the control that would have caught it. - Update the checklist edition and note the date. - Register risks: R-06. Severity: medium Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (read 2026-11-16). ## Related pages - [The full MCP Security Checklist](/mcp-security-checklist/) - [Observe](/mcp-security-checklist/observe/) - [Discover](/mcp-security-checklist/discover/) - [Method](/mcp-security-checklist/methodology/) - [MCP Risk Register](/mcp-risk-register/) --- # MCP security datasets URL: https://mcpservercybersecurity.com/datasets/ Updated: 2026-11-16 Two datasets are free to download as CSV and JSON under CC BY 4.0: the MCP Risk Register, with severity, impact, controls and sources for each row, and the MCP Security Checklist, ordered discover, vet, authenticate, constrain, observe and respond. Each download page carries the edition date and a citation line. ## Downloads - [MCP Risk Register](/datasets/mcp-risk-register/): [CSV](/datasets/mcp-risk-register.csv), [JSON](/datasets/mcp-risk-register.json) - [MCP Security Checklist](/datasets/mcp-security-checklist/): [CSV](/datasets/mcp-security-checklist.csv), [JSON](/datasets/mcp-security-checklist.json) ## License and citation CC BY 4.0. Cite assets as: MCP Server Security, "", edition 16 Nov 2026, https://mcpservercybersecurity.com//, licensed CC BY 4.0. ## Related pages - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) - [Method](/method/) - [The ranking](/ranking/) --- # MCP Risk Register dataset: CSV and JSON download URL: https://mcpservercybersecurity.com/datasets/mcp-risk-register/ Updated: 2026-11-16 The MCP Risk Register is published as a CSV file and a JSON file under CC BY 4.0. It holds 42 items across 6 categories, read on 2026-11-16. Cite it with the line below when you reuse the data, and link back to this page so readers can check the files against the sources. ## Files - [MCP Risk Register as CSV](/datasets/mcp-risk-register.csv) - [MCP Risk Register as JSON](/datasets/mcp-risk-register.json) ## How to cite Cite assets as: MCP Server Security, "", edition 16 Nov 2026, https://mcpservercybersecurity.com//, licensed CC BY 4.0. ## Related pages - [The MCP Risk Register](/mcp-risk-register/) - [Method](/mcp-risk-register/methodology/) - [All datasets](/datasets/) - [Scoring method](/method/) --- # MCP Security Checklist dataset: CSV and JSON download URL: https://mcpservercybersecurity.com/datasets/mcp-security-checklist/ Updated: 2026-11-16 The MCP Security Checklist is published as a CSV file and a JSON file under CC BY 4.0. It holds 42 items across 6 categories, read on 2026-11-16. Cite it with the line below when you reuse the data, and link back to this page so readers can check the files against the sources. ## Files - [MCP Security Checklist as CSV](/datasets/mcp-security-checklist.csv) - [MCP Security Checklist as JSON](/datasets/mcp-security-checklist.json) ## How to cite Cite assets as: MCP Server Security, "", edition 16 Nov 2026, https://mcpservercybersecurity.com//, licensed CC BY 4.0. ## Related pages - [The MCP Security Checklist](/mcp-security-checklist/) - [Method](/mcp-security-checklist/methodology/) - [All datasets](/datasets/) - [Scoring method](/method/) --- # MCP server security guides URL: https://mcpservercybersecurity.com/guides/ Updated: 2026-11-16 Twelve guides cover how to secure MCP servers, from the flagship best-practices guide to vetting, review, approved catalogs, scanning, language-specific hardening for Spring Boot, Python and TypeScript, cloud hosting, audit logging and incident response. Each claim is tied to the MCP specification, OWASP, advisories or vendor documentation read on 16 Nov 2026. ## 12 guides - [Building an Approved MCP Server Catalog for Your Company](/guides/building-an-approved-mcp-server-catalog/): How to run a company MCP server catalog: entry fields, approvers, enforcement in Claude Code and Cursor, and re-review triggers. - [How to Review an MCP Server for Security: Code and Tools](/guides/how-to-review-an-mcp-server-security/): A security review routine for an MCP server: read the tool list and code, test authorization, inputs and transport, and write a report. - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/): A pre-use vetting routine for MCP servers: publisher, source, startup command, tool descriptions and scopes, with the signals that mean stop. - [MCP Audit Logging: What to Record and How Long to Keep It](/guides/mcp-audit-logging-what-to-record/): Which fields an MCP audit log needs, what to redact, where to send logs and why retention must be checked, sourced to OWASP and vendor docs. - [MCP Security Best Practices for Server Operators](/guides/mcp-security-best-practices/): Eleven MCP security best practices for developer machines and remote endpoints, sourced to the MCP specification and OWASP, with a checklist map. - [MCP Security Incident Response: Steps for a Bad Server](/guides/mcp-security-incident-response-steps/): What to do when an MCP server is found malicious or vulnerable: contain, find every machine, revoke tokens and confirm cleanup, with sources. - [MCP Security vs API Security: What Changes](/guides/mcp-security-vs-api-security/): How MCP security differs from API security: model-chosen calls, shared tool descriptions and local servers, and which API controls still apply. - [MCP Server Scanning and Testing: Tools and Limits](/guides/mcp-server-scanning-and-testing/): Which scanners check MCP servers, what they miss, where to run them and how to test a server you build, with open-source options named. - [MCP Server Security on AWS, Azure and Google Cloud](/guides/mcp-server-security-on-aws-azure-and-google-cloud/): What AWS AgentCore Runtime, Azure API Management and Google Cloud Run provide for hosted MCP servers, and which settings stay with you. - [Securing a Python MCP Server: Host Checks and Deployment](/guides/securing-a-python-mcp-server/): Fix the Invalid Host header 421 error in the Python MCP SDK without switching off protection, and set transport security for deployment. - [Securing a Spring Boot MCP Server: Spring AI Settings](/guides/securing-a-spring-boot-mcp-server/): How to add OAuth 2.0 to a Spring AI MCP server, protect tools with method security and check which settings expose the endpoint. - [Securing a TypeScript MCP Server: Input, Binding and Tokens](/guides/securing-a-typescript-mcp-server/): Validate tool input with schemas, bind an Express MCP server to localhost with host checks and verify bearer tokens by audience in the TypeScript SDK. ## Related pages - [Glossary](/glossary/) - [The ranking](/ranking/) - [FAQ](/faq/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Building an approved MCP server catalog URL: https://mcpservercybersecurity.com/guides/building-an-approved-mcp-server-catalog/ Updated: 2026-11-16 An approved MCP server catalog is a maintained list of servers your company has reviewed, with owners, versions, allowed tools and a review date, enforced through the client's own allowlist settings. The public MCP Registry verifies namespaces but not code, so approval stays with you. Sources were read on 16 Nov 2026. ## How do you run an MCP registry for your company? Run it as a maintained list that points to reviewed servers, using the public registry's metadata format as a model. The MCP Registry stores a server.json with the server's unique name, where to locate it (a package name or remote URL), execution instructions and discovery data, and it says downstream aggregators can add security checks, ratings or review. A company catalog is that review layer. Keep it in version control so every addition is a reviewed change with an author and a date. Name entries consistently. The public registry uses reverse-DNS style names such as io.github.user/server-name, tied to a verified GitHub account or domain, and mirroring that convention makes it simple to map an internal entry to its public origin. Sources: [MCP Registry: About](https://modelcontextprotocol.io/registry/about). ## Who approves a server? A security reviewer approves a server, a named owner stays accountable for it and the requester never approves their own request. Splitting the roles keeps the catalog from becoming a rubber stamp. Vendors describe the same workflow in their products: Noma Security says it provides an MCP registry where servers are approved, reviewed or blocked, and Snyk documents an Allowed MCP Servers classification where unlisted servers show as Not Allowed. Keep the approval record in the same repository as the catalog. A pull request that names the reviewer, links the review report and shows the diff of tool definitions gives you the audit trail an auditor will ask for, and a request with no named owner should be refused, because an unowned server is the first one left unpatched. Sources: [Noma Security: MCP security](https://noma.security/solutions/mcp-security); [Snyk documentation: MCP Governance](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md). ## How do you enforce the catalog on developer machines? Enforce it through the client's administrator settings and, where you need more, an endpoint tool. Claude Code documents allowedMcpServers and deniedMcpServers settings and a managed-mcp.json file for exclusive control of the servers available, and Cursor documents an MCP Allowlist with command entries for local stdio servers, URL entries for remote servers and tool allowlists. Check the defaults. Snyk's documentation says its allowlist ships empty and disabled, and Cursor's tool allowlist allows all tools from a server when left empty. An unset list approves nothing and blocks nothing. Tool-level lists matter because one approved server can expose many tools. Cursor lets an administrator restrict which tools of an approved server run automatically, so approve read tools first and hold write and delete tools for a separate review. Sources: [Claude Code documentation: Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp); [Cursor documentation: Model Context Protocol](https://cursor.com/docs/context/mcp); [Snyk documentation: MCP Governance](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md). ## What should each catalog entry record? Each entry should record enough to repeat the review and to detect drift. Fields that earn their place are: - Server name, publisher and the namespace it was verified under. - Pinned version or content hash of the reviewed tool definitions. - Allowed tools and the scopes or credentials granted. - Transport and the exact command or URL. - Owner, approver and review date. - Next review trigger. Sources: [MCP Registry: About](https://modelcontextprotocol.io/registry/about); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How often should entries be re-reviewed? Re-review on every change to the tool definitions and on a fixed calendar interval you set. OWASP asks for monitoring of tool descriptions after installation and for consent to be requested again when definitions change, and an entry whose hash no longer matches should drop out of the allowed set until someone reviews the difference. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources - [MCP Registry: About](https://modelcontextprotocol.io/registry/about) (modelcontextprotocol.io), read 2026-11-16 - [Noma Security: MCP security](https://noma.security/solutions/mcp-security) (noma.security), read 2026-11-16 - [Snyk documentation: MCP Governance](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md) (docs.snyk.io), read 2026-11-16 - [Claude Code documentation: Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp) (code.claude.com), read 2026-11-16 - [Cursor documentation: Model Context Protocol](https://cursor.com/docs/context/mcp) (cursor.com), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 ## Related pages - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [MCP Security Best Practices for Server Operators](/guides/mcp-security-best-practices/) - [MCP Registry (glossary)](/glossary/mcp-registry/) - [Server manifest (glossary)](/glossary/server-manifest/) - [All guides](/guides/) - [Securing a TypeScript MCP Server: Input, Binding and Tokens](/guides/securing-a-typescript-mcp-server/) - [How to Review an MCP Server for Security: Code and Tools](/guides/how-to-review-an-mcp-server-security/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # How to review an MCP server for security URL: https://mcpservercybersecurity.com/guides/how-to-review-an-mcp-server-security/ Updated: 2026-11-16 Review an MCP server by listing every tool with its schema, reading the code behind each for command, path and URL handling, testing token checks and transport settings, and writing a report with a decision and a re-review trigger. The review tests what the specification and OWASP require. Sources were read on 16 Nov 2026. ## How do you read a server's code and tool list? Start from the tool list, then read the code behind each tool. A server that supports tools declares the capability and exposes each tool with a name, description, input schema and optional annotations, and may send a notification when the list changes. For each tool, trace three things: where input becomes a shell command, a file path or an outbound URL, where credentials are loaded, and what is written to logs. OWASP lists the failure families to look for: injection, path traversal and server-side request forgery. Read the dependency tree as well as the entry file. OWASP asks for scanning of server dependencies for known vulnerabilities, and a small server can pull in dozens of packages that run with the same privileges. Sources: [MCP specification: Tools (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/server/tools); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How do you review authorization on a remote server? Confirm that the server rejects requests without a valid token, rejects tokens issued for a different audience, and never forwards a client's token downstream. The authorization specification requires servers to validate that tokens were issued for them and to answer invalid or expired tokens with 401. If the server proxies a third-party API with a static client ID, read the confused deputy mitigations: per-client consent stored before the third-party flow, exact redirect URI matching, and state values that are single-use and short-lived. PKCE is required of clients. Sources: [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## How do you review input handling? Treat every argument as untrusted, because it originates from model output influenced by whatever context the model has read. OWASP asks for strict JSON Schema, with additionalProperties false and pattern constraints on strings, and for sanitizing against SQL, OS command and path traversal injection. Check outputs too. Tool output returns to the model and often feeds another tool, so unsanitized output can cause downstream request forgery or command injection. Any tool that fetches a URL from a model-supplied parameter needs an allowlist. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How do you review session and transport settings? Check that HTTP servers validate the Origin header, bind to localhost when local, and treat session IDs as identifiers rather than credentials. The transport specification requires servers to respond 403 to a present but invalid Origin and says local servers should bind to 127.0.0.1 rather than 0.0.0.0. The security best practices say servers must not use sessions for authentication, must use secure non-deterministic session IDs and should bind them to user-specific information. Sources: [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## What does a review report contain? A review report names the exact version reviewed, the findings and a decision someone can act on. Keep it to a page so it is read. - Server name, version or hash, publisher and transport. - Table of tools with schema, annotations and observed behavior. - Credentials, scopes and where tokens are stored. - Network destinations and file paths the server touches. - Findings with severity and the control that fixes each. - Decision: approve, restrict or reject, with the approver and date. - Trigger for the next review. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Tools (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/server/tools). ## When should you review again? Review again whenever the tool definitions or the code change. OWASP recommends pinning reviewed definitions by cryptographic hash and requiring review when the hash changes, and notes that a hash catches metadata changes but not changed behavior behind an unchanged definition, so version updates need a code diff as well. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources - [MCP specification: Tools (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/server/tools) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports) (modelcontextprotocol.io), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 ## Related pages - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [MCP Server Scanning and Testing: Tools and Limits](/guides/mcp-server-scanning-and-testing/) - [Tool schema (glossary)](/glossary/tool-schema/) - [Path traversal (glossary)](/glossary/path-traversal/) - [All guides](/guides/) - [Building an Approved MCP Server Catalog for Your Company](/guides/building-an-approved-mcp-server-catalog/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # How to vet an MCP server before use URL: https://mcpservercybersecurity.com/guides/how-to-vet-an-mcp-server-before-use/ Updated: 2026-11-16 Vet an MCP server before it runs by confirming the publisher, reading the source and startup command, inspecting every tool description and schema, checking requested scopes and pinning what you approved. Stop on look-alike package names, commands that touch sensitive paths, or descriptions that address the model. Sources were read on 16 Nov 2026. ## What do you check in a server before it runs? Check who published it, what it executes at startup, what each tool claims to do and which credentials it asks for. OWASP's supply chain guidance is to install servers only from trusted, verified sources, review source code and tool definitions first, verify package integrity with checksums or signing, and scan dependencies for known vulnerabilities. The MCP specification adds that a client must show the exact startup command, untruncated, and get explicit approval before a one-click local server runs. Apply the same rule to config files you receive from a teammate or a repository. Do the check on a disposable container or a clean machine, never on a laptop that holds production credentials. A server that is only being reviewed should not be able to read anything real. - Publisher identity and the exact package name. - Source code for command construction, file access and outbound requests. - The startup command with every argument. - Every tool name, description and input schema. - Requested OAuth scopes and where tokens are stored. - Dependencies and their known advisories. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## Which signals show a server is unsafe? Unsafe servers show up as look-alike names, risky startup commands and descriptions written for the model. OWASP calls typosquatting a common attack, with a package name one character from a popular one as the example, and says tool descriptions, parameter names and return schemas can all carry hidden instructions. The specification lists startup patterns a client should highlight: commands with sudo, deletion of files, network operations, or file access outside expected directories, and anything that reaches the home directory, SSH keys or system directories. A server that changes its tool definitions after approval, often called a rug pull, is a third signal, which is why OWASP asks for monitoring of tool descriptions after installation. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## How do you verify who published a server? Check the publisher against a source the server's author does not control. The MCP Registry requires publishers to prove ownership of a namespace through GitHub, DNS or HTTP challenges, which tells you who owns a name but nothing about the code. Some tools add publisher signals. Backslash Security's free MCP Server Security Hub, published 2025-06-16, shows GitHub publisher verification and an attack vector per server, and Snyk documents trust signals in its MCP Risk Score: popularity, official first-party status, URL reputation and maintainer identity. These are vendor statements about their own services. Sources: [MCP Registry: About](https://modelcontextprotocol.io/registry/about); [Backslash Security: MCP Server Security Hub](https://www.backslash.security/blog/mcp-server-security-hub); [Snyk documentation: Agent Supply Chain Security](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md). ## What do automated scanners add to a manual review? Scanners add speed and consistency on tool descriptions, prompts and packages, and they do not replace reading the code. Cisco's open-source MCP Scanner combines YARA rules and LLM-based analysis and scans a server's tools, prompts, resources and instructions, and Snyk says a local stdio server is matched against its catalog and assessed from its signature without starting it. Use scanner output as a prompt for review. OWASP warns that filters for instruction-like text cannot establish that the remaining text is safe. Sources: [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner); [Snyk documentation: Agent Supply Chain Security](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How do you decide to approve, restrict or reject? Approve a server only for the tools and scopes you reviewed, restrict the rest, and reject on any unresolved signal. Cursor's allowlist can limit which tools of an approved server run automatically, and Pillar Security says per-server tool allow-lists run in Monitor or Block mode, so a partial approval is possible in practice. Record the decision with the version or hash reviewed, then re-review on any change. OWASP asks clients to re-prompt for consent when tool definitions change. Sources: [Cursor documentation: Model Context Protocol](https://cursor.com/docs/context/mcp); [Pillar Security: agentic endpoint](https://www.pillar.security/platform/agentic-endpoint); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 - [MCP Registry: About](https://modelcontextprotocol.io/registry/about) (modelcontextprotocol.io), read 2026-11-16 - [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner) (github.com), read 2026-11-16 - [Snyk documentation: Agent Supply Chain Security](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md) (docs.snyk.io), read 2026-11-16 - [Backslash Security: MCP Server Security Hub](https://www.backslash.security/blog/mcp-server-security-hub) (backslash.security), read 2026-11-16 - [Cursor documentation: Model Context Protocol](https://cursor.com/docs/context/mcp) (cursor.com), read 2026-11-16 - [Pillar Security: agentic endpoint](https://www.pillar.security/platform/agentic-endpoint) (pillar.security), read 2026-11-16 ## Related pages - [How to Review an MCP Server for Security: Code and Tools](/guides/how-to-review-an-mcp-server-security/) - [Building an Approved MCP Server Catalog for Your Company](/guides/building-an-approved-mcp-server-catalog/) - [Typosquatted server (glossary)](/glossary/typosquatted-server/) - [Server pinning (glossary)](/glossary/server-pinning/) - [All guides](/guides/) - [MCP Audit Logging: What to Record and How Long to Keep It](/guides/mcp-audit-logging-what-to-record/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # MCP audit logging: what to record URL: https://mcpservercybersecurity.com/guides/mcp-audit-logging-what-to-record/ Updated: 2026-11-16 An MCP audit log should record every tool invocation with its full parameters, the user, the client, the server and version, the decision taken and a timestamp, with secrets redacted, and export it to a SIEM. No MCP document sets a retention period, so set one and check any vendor's default. Sources were read on 16 Nov 2026. ## Which fields belong in an MCP audit log? Log enough to answer who called which tool on which server with what arguments and what happened. OWASP's cheat sheet asks for all MCP tool invocations to be logged with full parameters, user context and timestamps. Add the context an investigator needs: - Timestamp and time zone. - User identity and the machine or workload. - Client application and version. - Server name, version or definition hash, and transport. - Tool name and full arguments. - Decision (allowed, blocked or asked) and who approved. - Result status and, where policy allows, a reference to the response. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How long should logs be kept? Keep them as long as your own policy or regulation requires; none found on 16 Nov 2026 in the MCP specification or the OWASP cheat sheet, which set no retention period. Products differ, so read each default. Snyk's documentation, for example, lists a fixed seven-day data retention for its Agent Behavior Governance, which means a team that needs a longer record must export the data. Write the retention period into policy for each log tier, for example raw parameters for a short period and summaries for longer, so the record you keep matches the questions an auditor will ask. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Snyk documentation: Agent Behavior Governance](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance.md); [Snyk documentation: MCP Governance](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md). ## What should you redact from MCP logs? Redact secrets and personal data before the record is stored. OWASP says logs should have secrets and PII removed, which matters because tool arguments often carry tokens, file contents and customer data. Keep tokens out of configuration and logs altogether: OWASP advises using the operating system's credential store rather than plaintext config files. Decide who may read the logs. Full parameters are sensitive by nature, so restrict access to the security team and write the redaction rules before the first record is stored, not after the first incident. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Where should MCP logs go? Send them to the system your analysts already use. OWASP recommends feeding MCP logs into a SIEM for anomaly detection. Tools document several routes: Cisco's DefenseClaw lists OTLP, Splunk, file and webhook destinations, CrowdStrike says Guardian exports agent telemetry into Falcon Next-Gen SIEM, and Harmonic Security's documentation describes full invocation logs with parameters and responses. Lasso Security says it exports audit trails of every MCP connection and tool call to a SIEM; Bay says events reach a SIEM in OCSF format. These are vendor statements. Test the pipeline with a deliberate blocked call and confirm it reaches the SIEM with every field intact. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Cisco DefenseClaw documentation: observability](https://cisco-ai-defense.github.io/defenseclaw/docs/observability/); [CrowdStrike: Falcon Guardian defines next generation of AI security](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/); [Harmonic Security documentation: What is the MCP Gateway](https://docs.harmonicsecurity.app/mcp-gateway/getting-started/what-is-mcp-gateway); [Lasso Security: MCP security use case](https://www.lasso.security/use-cases/mcp-security); [Bay: AI agent investigation](https://www.bay.io/ai-agent-investigation?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=guide). ## Which alerts are worth setting on MCP activity? Alert on change and on volume. OWASP lists new tools being called, admin-level queries and abnormal call frequency, and asks teams to monitor for credentials from one server appearing in calls to another. Add an alert for any tool definition whose hash differs from the reviewed one, since that is the rug pull signal. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Can you trust a log you did not build? Treat it as evidence to be judged, not as proof. OWASP points readers to a separate cheat sheet on verifying third-party agent execution evidence for exactly this case. When the log comes from a vendor, check what it establishes: whether the record is created before or after the action, who can edit it and whether you can export it in full. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [Snyk documentation: MCP Governance](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-behavior-governance/mcp-governance.md) (docs.snyk.io), read 2026-11-16 - [Cisco DefenseClaw documentation: observability](https://cisco-ai-defense.github.io/defenseclaw/docs/observability/) (cisco-ai-defense.github.io), read 2026-11-16 - [CrowdStrike: Falcon Guardian defines next generation of AI security](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/) (crowdstrike.com), read 2026-11-16 - [Harmonic Security documentation: What is the MCP Gateway](https://docs.harmonicsecurity.app/mcp-gateway/getting-started/what-is-mcp-gateway) (docs.harmonicsecurity.app), read 2026-11-16 - [Lasso Security: MCP security use case](https://www.lasso.security/use-cases/mcp-security) (lasso.security), read 2026-11-16 - [Bay: AI agent investigation](https://www.bay.io/ai-agent-investigation?utm_source=mcpservercybersecurity.com&utm_medium=referral&utm_campaign=guide) (bay.io), read 2026-11-16 ## Related pages - [MCP Security Incident Response: Steps for a Bad Server](/guides/mcp-security-incident-response-steps/) - [MCP Security Best Practices for Server Operators](/guides/mcp-security-best-practices/) - [Tool description (glossary)](/glossary/tool-description/) - [MCP session (glossary)](/glossary/mcp-session/) - [All guides](/guides/) - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # MCP security best practices URL: https://mcpservercybersecurity.com/guides/mcp-security-best-practices/ Updated: 2026-11-16 The best MCP security practices are to discover every server, vet each before first use, authenticate remote endpoints with tokens issued for your server, constrain files, network and scopes, log every tool call and keep a removal plan. Each practice below cites the MCP specification, OWASP or client documentation read on 16 Nov 2026. ## What are the best practices for MCP security? The best practices for MCP security are to run only servers you have reviewed, give each server the narrowest credentials and file access it needs, require authentication on every remote endpoint, keep a person in the loop for sensitive tool calls and log every call. The MCP specification and the OWASP MCP Security Cheat Sheet agree on these points, and OWASP tells operators to treat each server as an untrusted, separate security domain. Eleven practices follow from those two documents: - Inventory every configured server on every machine, including project-level config files. - Read the source, tool descriptions and schemas before first use, and verify the package name against its publisher. - Show the exact startup command to the user before a local server runs. - Sandbox local servers with minimal file system, network and credential access. - Require authentication on any remote endpoint that exposes non-public tools or data. - Accept only tokens issued for your server, and never pass them to an upstream API. - Request narrow scopes and use short-lived, per-server credentials. - Validate every tool input as untrusted, and keep a human approval step for destructive calls. - Pin the complete reviewed tool definition and re-review when it changes. - Log each invocation with parameters, user and timestamp, and send logs to a SIEM. - Know in advance how to remove a server and revoke what it held. Sources: [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Which practices matter most for developer machines? Controlling what starts on the machine matters most, because a local MCP server is a program the client launches from a config entry. The specification lists three local attacks: a malicious startup command in a client configuration, a malicious payload inside the server itself, and an insecure local server reached through DNS rebinding. Its stated risks include arbitrary code execution with the client's privileges and no visibility into the commands run. Configuration files that ride along in a repository are a delivery path, since Visual Studio Code documents a workspace-root .mcp.json that teams share through source control. Review changes to those files like code. Clients also give administrators allowlists: Claude Code documents allowedMcpServers and deniedMcpServers settings and a managed-mcp.json file for exclusive control, and Cursor documents an MCP Allowlist that approves local servers by command pattern and remote servers by URL pattern. Cursor notes that allowlisting approves a configuration and does not install the server. Start with the inventory, because every other control depends on it. A team that cannot list its servers cannot allowlist, review or revoke them. Project-level files such as .mcp.json travel with the repository, user-level files sit in the home directory, and desktop apps keep their own configuration, so a complete inventory reads all three places on every operating system you support. Sources: [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices); [Visual Studio Code documentation: Add and use MCP servers](https://code.visualstudio.com/docs/copilot/customization/mcp-servers); [Claude Code documentation: Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp); [Cursor documentation: Model Context Protocol](https://cursor.com/docs/context/mcp). ## How do you secure the connection to a remote MCP server? Require authentication, validate the token's audience on every request, and refuse any token that was not issued for your server. The specification makes authorization optional, so a remote server with non-public tools has no protection until you add it. When OAuth is used, clients must send the resource parameter that names the server, and servers must reject tokens issued for another audience with a 401. Token passthrough, where a server accepts a client's token and forwards it downstream, is forbidden by the authorization specification. It defeats rate limiting and audit trails and widens the blast radius of a stolen token. For HTTP servers, validate the Origin header and answer an invalid one with 403, and bind local servers to 127.0.0.1 rather than all interfaces. Store tokens in the operating system's credential store, not in plaintext config files. Sources: [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices); [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How do you control what a tool can do? Review the whole tool definition, restrict what each tool can reach, and require approval for destructive calls. OWASP treats the entire schema as an injection surface, not only the description field, and advises hashing the complete definition, including input schema, output schema and annotations, so a changed definition triggers review. A hash detects metadata changes; it does not detect changed behavior behind an unchanged definition. The specification says clients must treat tool annotations as untrusted unless they come from trusted servers, and that there should always be a human in the loop who can deny tool invocations. Validate inputs as untrusted because they come from model output shaped by whatever the model has read. Use strict JSON Schema with additionalProperties set to false, and never fetch arbitrary URLs a model supplies without an allowlist. Match the control to the effect of the call. OWASP says a read or write label alone may not establish risk, and it recommends choosing between explicit confirmation, narrowly scoped authorization and policy checks enforced at execution time according to what the tool can do, including sensitive reads and irreversible actions. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Tools (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/server/tools). ## How do the practices map to the checklist? The checklist groups the practices into six steps, in the order a team usually works. Discover covers the inventory of servers and config files. Vet covers source, descriptions, publisher and provenance. Authenticate covers tokens, audience and scopes. Constrain covers sandboxing, allowlists and approval. Observe covers logging and alerting. Respond covers removal, revocation and finding every machine that ran a server. Each checklist item links back to the register row it reduces, so a ticked box can be traced to a risk. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## What do the standards leave to you? The standards leave code review, enforcement and monitoring to you. The MCP Registry verifies who owns a namespace and hosts metadata, but it delegates security scanning of server code to package registries and downstream aggregators. Developer tools carry the same risk as production servers. In June 2025 a critical advisory for MCP Inspector (CVE-2025-49596) described remote code execution in versions below 0.14.1 because the client and proxy did not authenticate each other, and the fix was to upgrade to 0.14.1 or later. Keep inspection tools updated and off shared networks. Message signing is optional too. OWASP describes an Internet-Draft for signing and replay protection as an individual work in progress and not an adopted MCP standard, and advises against presenting it as a requirement for ordinary deployments. Sources: [MCP Registry: About](https://modelcontextprotocol.io/registry/about); [GitHub advisory GHSA-7f8r-222p-6f5g: Inspector proxy server vulnerabilities (CVE-2025-49596)](https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-7f8r-222p-6f5g). ## Sources - [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Tools (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/server/tools) (modelcontextprotocol.io), read 2026-11-16 - [MCP Registry: About](https://modelcontextprotocol.io/registry/about) (modelcontextprotocol.io), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [GitHub advisory GHSA-7f8r-222p-6f5g: Inspector proxy server vulnerabilities (CVE-2025-49596)](https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-7f8r-222p-6f5g) (github.com), read 2026-11-16 - [Claude Code documentation: Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp) (code.claude.com), read 2026-11-16 - [Cursor documentation: Model Context Protocol](https://cursor.com/docs/context/mcp) (cursor.com), read 2026-11-16 - [Visual Studio Code documentation: Add and use MCP servers](https://code.visualstudio.com/docs/copilot/customization/mcp-servers) (code.visualstudio.com), read 2026-11-16 ## Related pages - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [MCP Audit Logging: What to Record and How Long to Keep It](/guides/mcp-audit-logging-what-to-record/) - [Building an Approved MCP Server Catalog for Your Company](/guides/building-an-approved-mcp-server-catalog/) - [MCP authorization (glossary)](/glossary/mcp-authorization/) - [Tool annotations (glossary)](/glossary/tool-annotations/) - [All guides](/guides/) - [MCP Security Incident Response: Steps for a Bad Server](/guides/mcp-security-incident-response-steps/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # MCP security incident response steps URL: https://mcpservercybersecurity.com/guides/mcp-security-incident-response-steps/ Updated: 2026-11-16 When an MCP server is found malicious, block it centrally, find every machine and config file that references it, assume anything the client could read is exposed, rotate and revoke credentials at the issuer, then confirm removal and record the gap. Claude Code's remove command does not revoke tokens by itself. Sources were read on 16 Nov 2026. ## What do you do when a server is found malicious? Contain it first, centrally if you can. Add the server to the deny side of the client's administrator settings: Claude Code documents deniedMcpServers, and Cursor's allowlist approves by command and URL pattern, so removing the pattern stops new use. Then preserve evidence before cleanup: the config entry, the package copy and the logs. The MCP specification says local servers run with the client's privileges, with arbitrary code execution, data exfiltration and data loss among the stated risks, so treat the host as exposed. Sources: [Claude Code documentation: Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp); [Cursor documentation: Model Context Protocol](https://cursor.com/docs/context/mcp); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## How do you find every machine that ran it? Search for the server's name, package or URL in client configuration files and in your inventory data. Microsoft says Defender for Endpoint's local agent inventory lists the remote and local MCP servers per agent and that the data can be queried in advanced hunting, and Snyk's Evo documentation says its fleet view lists every machine that reported a scan with counts of MCP servers discovered. Without such a tool, an MDM or EDR query over the known config locations does the same job. Cisco's open-source scanner has a mode that reads well-known client config locations, which is useful for a spot check. Sources: [Microsoft Learn: Discover local AI agents in Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents); [Snyk documentation: Agent Supply Chain Security](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md); [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner). ## How do you revoke what the server held? Revoke tokens at the issuer and rotate any key the server could read. Claude Code's documentation says that removing a server deletes the OAuth tokens and client registration it stored locally but does not revoke them at the authorization server, and that running claude mcp logout before claude mcp remove revokes them where the server supports revocation. Environment variables in config files, cloud keys and tokens in the credential store all count as readable by a server that ran with the user's rights. OWASP advises per-server credentials, which limits how many you rotate. Sources: [Claude Code documentation: Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## What if the server is vulnerable rather than malicious? Patch or retire it on a deadline and treat the exposure window like a compromise if it was reachable. The MCP Inspector advisory (CVE-2025-49596, critical) shows the pattern: versions below 0.14.1 let unauthenticated requests launch MCP commands over stdio, and the fix was to upgrade to 0.14.1 or later. Check where such tools were run, since a developer tool left listening on a shared network is reachable by others. Sources: [GitHub advisory GHSA-7f8r-222p-6f5g: Inspector proxy server vulnerabilities (CVE-2025-49596)](https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-7f8r-222p-6f5g). ## How do you confirm the cleanup worked? Re-scan the machines, check that the config entry is gone and look in the logs for any later call to the server. Confirm at the identity provider that the tokens no longer work. Close the incident by checking that the deny entry is in place centrally, so a returning config file does not bring the server back. Tell developers what happened and ask them to report any server they added themselves, since the inventory only knows what it found. Sources: [Claude Code documentation: Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## What do you record afterwards? Record the timeline, the machines and credentials affected, how the server was approved and which control would have stopped it. Feed the answer into the catalog and the checklist, for example by adding hash pinning or a stricter startup-command review. OWASP's recommendation of regular audits and simulated attacks against MCP setups gives a way to test the fix. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources - [Claude Code documentation: Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp) (code.claude.com), read 2026-11-16 - [Cursor documentation: Model Context Protocol](https://cursor.com/docs/context/mcp) (cursor.com), read 2026-11-16 - [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 - [Microsoft Learn: Discover local AI agents in Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents) (learn.microsoft.com), read 2026-11-16 - [Snyk documentation: Agent Supply Chain Security](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md) (docs.snyk.io), read 2026-11-16 - [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner) (github.com), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [GitHub advisory GHSA-7f8r-222p-6f5g: Inspector proxy server vulnerabilities (CVE-2025-49596)](https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-7f8r-222p-6f5g) (github.com), read 2026-11-16 ## Related pages - [MCP Audit Logging: What to Record and How Long to Keep It](/guides/mcp-audit-logging-what-to-record/) - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [MCP config file (glossary)](/glossary/mcp-config-file/) - [npx launcher (glossary)](/glossary/npx-launcher/) - [All guides](/guides/) - [MCP Security Best Practices for Server Operators](/guides/mcp-security-best-practices/) - [MCP Security vs API Security: What Changes](/guides/mcp-security-vs-api-security/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # MCP security vs API security URL: https://mcpservercybersecurity.com/guides/mcp-security-vs-api-security/ Updated: 2026-11-16 MCP security differs from API security because a model decides which tool to call and with what arguments, every connected server's tool descriptions share one context, and local servers run code on the machine. Authentication, TLS, rate limits, input validation and logging carry over. Sources were read on 16 Nov 2026. ## How is MCP security different from API security? MCP security differs because the caller is a model choosing actions at run time. OWASP's cheat sheet puts it directly: in traditional APIs developers control every call, while MCP lets language models decide which tools to invoke, when and with which parameters. Two consequences follow. Arguments are untrusted even from your own client, since they come from model output shaped by whatever the model has read. And the model sees the tool descriptions of all connected servers at once, which is what makes cross-server attacks possible. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Which API controls carry over? Most of the transport and identity controls carry over unchanged. Require authentication on remote endpoints, validate tokens for the right audience, use TLS for remote connections, apply rate limits, quotas and timeouts per session or tenant, and validate every input. The MCP authorization specification builds on OAuth 2.1 with resource indicators and protected resource metadata, so existing identity providers do most of the work. Platform tools carry over as well: Azure API Management documents key-based and Microsoft Entra token authentication for MCP servers it manages. Limits need a different shape for a model that can loop. Count calls per session and per tool rather than per address, and for paid integrations OWASP asks for a cap on total spending per period across all servers, with authorized but unsettled payments counted at their maximum. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization); [Microsoft Learn: Secure access to MCP servers in Azure API Management](https://learn.microsoft.com/en-us/azure/api-management/secure-mcp-servers). ## Which risks are new in MCP? Tool poisoning, rug pulls, tool shadowing and local server code execution have no direct counterpart in a conventional API. OWASP defines tool poisoning as malicious instructions hidden in tool descriptions, schemas or return values, a rug pull as a server changing its definitions after approval, and shadowing as one server's description manipulating the agent's use of another's tools. Local servers add a risk of their own. The specification lists a malicious startup command in a client configuration and warns that servers run with the same privileges as the client. Shared context makes these risks compound. One poisoned server can influence calls to a trusted one, which is why OWASP says to treat each server as a separate security domain and to monitor cross-server data flows, such as credentials from one server appearing in calls to another. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## What changes for authorization and consent? Authorization has to cover the person approving the action, not only the token. The specification states that hosts must obtain explicit user consent before invoking a tool, and that there should always be a human in the loop who can deny invocations. Token handling is stricter than many APIs apply: the server must not accept a token that was not issued for it, and token passthrough to a downstream API is forbidden. OWASP also advises treating tool annotations as hints rather than enforcement. Sources: [MCP specification: Tools (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/server/tools); [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## What changes for logging and monitoring? Log the model-driven call, not only the HTTP request. OWASP asks for every tool invocation to be logged with full parameters, user context and timestamps, plus alerts on new tools being called, admin-level queries and unusual call frequency. Gateways that sit in front of servers form a separate category that this site does not cover. Include the hash of the tool description in the record, so a later investigation can tell which version of the tool the model saw. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Tools (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/server/tools) (modelcontextprotocol.io), read 2026-11-16 - [Microsoft Learn: Secure access to MCP servers in Azure API Management](https://learn.microsoft.com/en-us/azure/api-management/secure-mcp-servers) (learn.microsoft.com), read 2026-11-16 ## Related pages - [MCP Security Best Practices for Server Operators](/guides/mcp-security-best-practices/) - [MCP Audit Logging: What to Record and How Long to Keep It](/guides/mcp-audit-logging-what-to-record/) - [Tool shadowing (glossary)](/glossary/tool-shadowing/) - [Tool description (glossary)](/glossary/tool-description/) - [All guides](/guides/) - [MCP Security Incident Response: Steps for a Bad Server](/guides/mcp-security-incident-response-steps/) - [MCP Server Scanning and Testing: Tools and Limits](/guides/mcp-server-scanning-and-testing/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # MCP server scanning and testing URL: https://mcpservercybersecurity.com/guides/mcp-server-scanning-and-testing/ Updated: 2026-11-16 Scanners that check MCP servers include Cisco's MCP Scanner and Snyk's Agent Scan, both Apache-2.0 open source, plus OWASP's named mcp-scan. They read tool metadata and packages, so they miss behavior hidden behind unchanged definitions. Test your own servers for token, Origin and input handling. Sources were read on 16 Nov 2026. ## Which scanners check MCP servers? Several open-source and vendor scanners check MCP servers, and two documented ones carry Apache-2.0 licenses. Cisco's MCP Scanner scans a server's tools, prompts, resources and instructions with YARA rules and LLM-based analysis, and has a mode that reads servers from well-known client config locations on the machine. Snyk's agent-scan README lists a per-operating-system table of clients it reads, including Claude Desktop, Claude Code, Cursor and VS Code. OWASP's cheat sheet names mcp-scan for detecting poisoned descriptions and cross-server shadowing, and Backslash Security runs a free MCP Server Security Hub that scores servers without a login. Each tool answers a different question. Cisco's scanner reads server contents, Snyk's reads client configs on a machine to find servers, and a hub score summarizes a public server. Pick by the question you need answered, and expect to combine two. Sources: [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner); [Snyk agent-scan README](https://raw.githubusercontent.com/snyk/agent-scan/main/README.md); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Backslash Security: MCP Server Security Hub](https://www.backslash.security/blog/mcp-server-security-hub). ## What do scanners miss? Scanners miss behavior that is not visible in the metadata or code they read. OWASP states that hashing a tool definition detects metadata changes, not changes to server code or behavior behind an unchanged definition, and that filters for injected instructions cannot establish that the remaining text is safe. Coverage gaps matter too. The MCP Registry delegates code scanning to others, and Snyk's supply chain documentation admits that a stdio server not yet in its catalog ends as discovery only, with no risk assessment. Ask any scanner what it does with a server it has never seen. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP Registry: About](https://modelcontextprotocol.io/registry/about); [Snyk documentation: Agent Supply Chain Security](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md). ## Where do you run scans? Run them at three points: when a server is proposed, on every laptop on a schedule, and in CI for servers you build. Proposal-time scans feed the catalog review, and scheduled laptop scans catch servers added outside the process. Snyk's deployment documentation says its scheduled discovery scan runs every 24 hours by default on macOS and Windows, with the installer pushed by Jamf Pro or Intune. Cisco's MCP Scanner is a command-line tool, and its DefenseClaw documentation describes an enterprise edition deployed with a package or an MDM. Pick the cadence that matches how fast your developers add servers. Sources: [Snyk documentation: Activation and deployment](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/activation-and-deployment.md); [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner); [Cisco DefenseClaw documentation: support matrix](https://cisco-ai-defense.github.io/defenseclaw/docs/support-matrix/). ## How do you test a server you build? Test it with an MCP client you control, and write automated checks for the behaviors the specification requires. The MCP Inspector is the usual manual client, and its own advisory is a reminder to keep it patched: versions below 0.14.1 allowed unauthenticated requests to launch MCP commands over stdio (CVE-2025-49596). Automate these expectations in CI: - A request with no token, an expired token or a token for another audience returns 401. - A request carrying an invalid Origin header returns 403. - A tool call with an undeclared parameter or a malformed value is rejected by the schema. - A tool that fetches a URL refuses internal addresses and cloud metadata endpoints. Sources: [GitHub advisory GHSA-7f8r-222p-6f5g: Inspector proxy server vulnerabilities (CVE-2025-49596)](https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-7f8r-222p-6f5g); [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization); [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How do you act on scan results? Treat a finding as a reason to review, not as a verdict. Map each finding to a register row, decide whether the control already exists, and record the outcome in the catalog entry. A clean scan supports approval but does not replace the code review for a server with file or shell access. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources - [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner) (github.com), read 2026-11-16 - [Cisco DefenseClaw documentation: support matrix](https://cisco-ai-defense.github.io/defenseclaw/docs/support-matrix/) (cisco-ai-defense.github.io), read 2026-11-16 - [Snyk agent-scan README](https://raw.githubusercontent.com/snyk/agent-scan/main/README.md) (raw.githubusercontent.com), read 2026-11-16 - [Snyk documentation: Agent Supply Chain Security](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/agent-supply-chain-security.md) (docs.snyk.io), read 2026-11-16 - [Snyk documentation: Activation and deployment](https://docs.snyk.io/agent-security/evo-by-snyk/agentic-development-security-ads/activation-and-deployment.md) (docs.snyk.io), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [MCP Registry: About](https://modelcontextprotocol.io/registry/about) (modelcontextprotocol.io), read 2026-11-16 - [Backslash Security: MCP Server Security Hub](https://www.backslash.security/blog/mcp-server-security-hub) (backslash.security), read 2026-11-16 - [GitHub advisory GHSA-7f8r-222p-6f5g: Inspector proxy server vulnerabilities (CVE-2025-49596)](https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-7f8r-222p-6f5g) (github.com), read 2026-11-16 - [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [How to Review an MCP Server for Security: Code and Tools](/guides/how-to-review-an-mcp-server-security/) - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [MCP Inspector (glossary)](/glossary/mcp-inspector/) - [Tool shadowing (glossary)](/glossary/tool-shadowing/) - [All guides](/guides/) - [MCP Security vs API Security: What Changes](/guides/mcp-security-vs-api-security/) - [MCP Server Security on AWS, Azure and Google Cloud](/guides/mcp-server-security-on-aws-azure-and-google-cloud/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # MCP server security on AWS, Azure and Google Cloud URL: https://mcpservercybersecurity.com/guides/mcp-server-security-on-aws-azure-and-google-cloud/ Updated: 2026-11-16 AWS, Azure and Google Cloud each document a way to put authentication in front of a hosted MCP server: a JWT authorizer on AgentCore Runtime, key and Entra ID token validation in API Management, and IAM invoker access on Cloud Run. Audience checks, scopes, egress and logging remain the customer's job. Sources were read on 16 Nov 2026. ## What do cloud providers offer for hosted MCP servers? Each provider documents an authentication layer in front of the server. On AWS, the AgentCore Runtime guide deploys an MCP server and its example configures a CUSTOM_JWT authorizer with an OpenID discovery URL and a list of allowed clients, and passes the Authorization header through to the runtime. On Azure, API Management can expose and govern MCP servers and secure inbound access (client to API Management) and outbound access (API Management to the server). Inbound options include a subscription key and OAuth 2.1 tokens issued by Microsoft Entra ID, validated with the validate-azure-ad-token policy. On Google Cloud, a Cloud Run service requires the Cloud Run Invoker IAM role by default, and the documentation lists IAM invoker permission, OIDC ID tokens and Identity-Aware Proxy as ways to authenticate clients. Sources: [AWS documentation: Deploy MCP servers in AgentCore Runtime](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-mcp.html); [Microsoft Learn: Secure access to MCP servers in Azure API Management](https://learn.microsoft.com/en-us/azure/api-management/secure-mcp-servers); [Google Cloud documentation: Host MCP servers on Cloud Run](https://cloud.google.com/run/docs/host-mcp-servers). ## Which settings are the customer's job? The audience, the scopes, the network path and the logs stay with you. The MCP authorization specification requires the server to validate that each token was issued for it, so confirm that the provider's authorizer checks the audience you intend, and keep a check in server code as defense in depth. Scopes are yours too. OWASP advises narrow, per-server credentials and short-lived tokens, and Azure's guidance on subscription keys is to store them in workspace settings or secure inputs rather than in the configuration file. Network exposure is also a choice you make. A Cloud Run service can be opened to the public, and API Management can front an MCP server hosted outside it, so confirm which path a client actually takes before you describe the server as protected. Sources: [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [Microsoft Learn: Secure access to MCP servers in Azure API Management](https://learn.microsoft.com/en-us/azure/api-management/secure-mcp-servers). ## How do local clients authenticate to a remote server? Use the platform identity the provider documents and avoid sharing static secrets. Google's documentation shows running a local Cloud Run proxy so the MCP client reaches the remote service using your own credentials and the invoker role, rather than opening the service to the public. For clients that must use OAuth, follow the authorization specification: the client sends the resource parameter, and the server answers an invalid or expired token with 401. Rotate any shared secret on a schedule and prefer short-lived tokens, as OWASP advises over long-lived personal access tokens. Sources: [Google Cloud documentation: Host MCP servers on Cloud Run](https://cloud.google.com/run/docs/host-mcp-servers); [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization). ## What should you check before a hosted server is reachable? Check that authentication is on, that the server cannot be tricked into reaching internal addresses and that it validates Origin. The MCP security best practices describe server-side request forgery against addresses such as the 169.254.169.254 cloud metadata endpoint and advise blocking private and link-local ranges, requiring HTTPS and using an egress proxy for server-side deployments. Avoid writing the IP validation yourself, because custom parsers miss encoded forms; use a maintained library or an egress proxy. Sources: [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices); [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports). ## What should a hosted server log? Log every tool invocation with parameters, the caller's identity and a timestamp, and send the records to your SIEM. OWASP also asks for redaction of secrets and personal data before storage. Log at the MCP server, not only at the cloud front door: the specification notes that downstream logs can show a different identity than the real caller when tokens are passed through. Cloud audit logs show who reached the service and not which tool ran with which arguments, so the application log is still needed. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## Sources - [AWS documentation: Deploy MCP servers in AgentCore Runtime](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-mcp.html) (docs.aws.amazon.com), read 2026-11-16 - [Microsoft Learn: Secure access to MCP servers in Azure API Management](https://learn.microsoft.com/en-us/azure/api-management/secure-mcp-servers) (learn.microsoft.com), read 2026-11-16 - [Google Cloud documentation: Host MCP servers on Cloud Run](https://cloud.google.com/run/docs/host-mcp-servers) (cloud.google.com), read 2026-11-16 - [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [MCP Security vs API Security: What Changes](/guides/mcp-security-vs-api-security/) - [MCP Audit Logging: What to Record and How Long to Keep It](/guides/mcp-audit-logging-what-to-record/) - [SSRF (glossary)](/glossary/ssrf/) - [Resource indicators (glossary)](/glossary/resource-indicators/) - [All guides](/guides/) - [MCP Server Scanning and Testing: Tools and Limits](/guides/mcp-server-scanning-and-testing/) - [Securing a Python MCP Server: Host Checks and Deployment](/guides/securing-a-python-mcp-server/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Securing a Python MCP server URL: https://mcpservercybersecurity.com/guides/securing-a-python-mcp-server/ Updated: 2026-11-16 In the Python MCP SDK, a streamable HTTP server accepts only localhost Host headers unless you pass transport_security with your real hostnames. A 421 Invalid Host header means that check fired. Add allowed hosts and origins instead of disabling it, and configure proxy headers behind TLS termination. Sources were read on 16 Nov 2026. ## What does the invalid host header error mean? It means the server's DNS-rebinding protection rejected a request whose Host header it did not expect. In the Python SDK, a streamable HTTP app created with no transport_security setting accepts a request only if its Host header is 127.0.0.1, localhost or [::1] with a port, and it answers anything else with 421 Misdirected Request and the body Invalid Host header. The Python client does not show those words. It reports MCPError: Server returned an error response, and the reason appears only in the server log as a warning from mcp.server.transport_security. The error is most common right after a first deployment behind a real hostname. After a deployment, request the public URL using the hostname you configured and read the server log. A warning for a hostname you do not recognize means a request arrived with a Host header you did not plan for, which is worth tracing before you widen the list. Sources: [MCP Python SDK documentation: Troubleshooting](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/troubleshooting.md); [MCP Python SDK documentation: Deploy and scale](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/run/deploy.md). ## How do you turn on transport security? Transport security is already on by default, so the task is to tell it which hostnames are yours. Pass transport_security=TransportSecuritySettings(allowed_hosts=[...], allowed_origins=[...]) with the names you actually serve. The SDK documentation gives three details. Entries in allowed_hosts are exact strings, so mcp.example.com matches a bare Host header and mcp.example.com:* matches any port, and you should list both. The allowed_origins list matters only for browsers, because nothing else sends an Origin header. Switching the check off is described as the honest configuration only behind a reverse proxy that already controls the Host header. List only production names. A wildcard or a long list copied from a test environment widens the set of Host headers the server accepts, which is the exposure the check exists to limit. Sources: [MCP Python SDK documentation: Deploy and scale](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/run/deploy.md); [MCP Python SDK documentation: Troubleshooting](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/troubleshooting.md). ## Why does the host check matter on a laptop? It stops a malicious web page from driving your local server through a DNS name that was rebound to 127.0.0.1. The MCP transport specification says servers must validate the Origin header on all incoming connections, must respond 403 when it is present and invalid, and should bind only to localhost rather than all interfaces. Local convenience is the usual reason the check gets removed. Keep it on for any server that holds file or credential access. Sources: [MCP Python SDK documentation: Deploy and scale](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/run/deploy.md); [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports). ## Which deployment settings change behind a proxy? Run uvicorn with proxy headers enabled and a list of the proxy addresses you trust. The SDK documentation says that when TLS ends at an ingress, load balancer or reverse proxy and uvicorn serves plain HTTP behind it, uvicorn must trust the proxy's X-Forwarded headers, or its redirects point at http:// and the client refuses them. The same page warns that the wildcard value for forwarded-allow-ips trusts every hop and is right only when nothing but the proxy can reach uvicorn. Restrict network access to the server port accordingly. Sources: [MCP Python SDK documentation: Deploy and scale](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/run/deploy.md). ## What else does the SDK documentation warn about? Duplicate tool names resolve silently: when two registrations use the same name, the first one wins, the second is dropped and a warning in the server log is the only signal. Keep that warning on and read the log after startup, since a tool you thought was live may be a different implementation. Transport security is not authentication. A server with non-public tools still needs token checks that validate the audience, and every tool input is untrusted, as OWASP's cheat sheet puts it. Sources: [MCP Python SDK documentation: Troubleshooting](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/troubleshooting.md); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization). ## Sources - [MCP Python SDK documentation: Troubleshooting](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/troubleshooting.md) (raw.githubusercontent.com), read 2026-11-16 - [MCP Python SDK documentation: Deploy and scale](https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/docs/run/deploy.md) (raw.githubusercontent.com), read 2026-11-16 - [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports) (modelcontextprotocol.io), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Securing a TypeScript MCP Server: Input, Binding and Tokens](/guides/securing-a-typescript-mcp-server/) - [Securing a Spring Boot MCP Server: Spring AI Settings](/guides/securing-a-spring-boot-mcp-server/) - [DNS rebinding (glossary)](/glossary/dns-rebinding/) - [Localhost binding (glossary)](/glossary/localhost-binding/) - [All guides](/guides/) - [MCP Server Security on AWS, Azure and Google Cloud](/guides/mcp-server-security-on-aws-azure-and-google-cloud/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Securing a Spring Boot MCP server URL: https://mcpservercybersecurity.com/guides/securing-a-spring-boot-mcp-server/ Updated: 2026-11-16 Secure a Spring AI MCP server by adding the community mcp-server-security module with Spring Security, configuring the OAuth 2.0 resource server with your issuer URL, and protecting individual tools with method security. Check the protocol, endpoint path and any permit-all matchers before deploying. Spring documents the module as community-driven and work in progress. ## How do you add authentication to a Spring AI MCP server? Add the mcp-server-security module, Spring Security and, for OAuth 2.0, the OAuth2 resource server starter, then configure the MCP resource server with your authorization server's issuer URI. The Spring AI reference lists org.springaicommunity:mcp-server-security, spring-boot-starter-security and spring-boot-starter-oauth2-resource-server as the dependencies, and shows a security filter chain that applies the MCP OAuth configurer with the issuer URL as the required setting. The module also supports API key authentication in a basic form. Prefer OAuth for anything beyond a single trusted caller, because the MCP authorization specification is built on OAuth 2.1. Sources: [Spring AI reference: MCP Security](https://docs.spring.io/spring-ai/reference/api/mcp/mcp-security.html); [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization). ## Which settings expose the server? The protocol, the endpoint path and the request matchers decide what is reachable. The server starter selects the transport with spring.ai.mcp.server.protocol, where STREAMABLE enables streamable HTTP, and the streamable endpoint defaults to /mcp. Read every matcher in your filter chain. A permitAll rule on the MCP path in a copied sample opens the tools to anyone who can reach it, and the OWASP guidance is to require authentication whenever a remote endpoint exposes non-public tools or data. The reference also notes that using MCP Inspector may need CSRF and CORS protection disabled; treat that as a development setting and keep it out of deployed configuration. If the server is meant only for local clients, keep it off public interfaces: the MCP transport specification says local servers should bind only to localhost and must validate the Origin header on HTTP connections. Sources: [Spring AI reference: Streamable-HTTP MCP Servers](https://docs.spring.io/spring-ai/reference/api/mcp/mcp-streamable-http-server-boot-starter-docs.html); [Spring AI reference: MCP Security](https://docs.spring.io/spring-ai/reference/api/mcp/mcp-security.html); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports). ## How do you protect individual tools? Annotate tool methods with Spring Security's method security so each tool checks the caller. The reference shows an @PreAuthorize annotation on an @McpTool method with isAuthenticated() as the expression. Go further than authenticated: use authorities or scopes per tool, so a read tool and a delete tool need different grants. This matches the specification's scope minimization guidance and OWASP's advice to request narrow scopes. Test the result with a token that is missing the required authority. The call should fail with an authorization error and the tool body should never run, and that test belongs in the build so a refactor cannot remove an annotation unnoticed. Sources: [Spring AI reference: MCP Security](https://docs.spring.io/spring-ai/reference/api/mcp/mcp-security.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## What does the module leave to you? The module covers authentication and tool authorization; it does not review your tool code. Validate every tool argument as untrusted, avoid building shell commands or file paths from raw input, and restrict outbound URLs. Spring documents the module as community-driven and not yet endorsed by Spring AI or the MCP project, compatible with WebMVC-based servers only, so check the project repository for the Spring AI version it supports before upgrading. Keep the authorization server separate from the MCP server where you can. The module family includes an authorization server component, but a shared identity provider with audit trails and revocation is easier to operate and review. Sources: [Spring AI reference: MCP Security](https://docs.spring.io/spring-ai/reference/api/mcp/mcp-security.html); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How do you disable operations you do not need? Turn off what the deployment does not use. The streamable-HTTP properties include a disallow-delete setting, which defaults to false, and the stateless protocol option avoids server-held sessions entirely. The MCP security best practices say servers must not use sessions for authentication, so whichever mode you pick, verify the token on each request. Sources: [Spring AI reference: Streamable-HTTP MCP Servers](https://docs.spring.io/spring-ai/reference/api/mcp/mcp-streamable-http-server-boot-starter-docs.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## Sources - [Spring AI reference: MCP Security](https://docs.spring.io/spring-ai/reference/api/mcp/mcp-security.html) (docs.spring.io), read 2026-11-16 - [Spring AI reference: Streamable-HTTP MCP Servers](https://docs.spring.io/spring-ai/reference/api/mcp/mcp-streamable-http-server-boot-starter-docs.html) (docs.spring.io), read 2026-11-16 - [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Securing a Python MCP Server: Host Checks and Deployment](/guides/securing-a-python-mcp-server/) - [Securing a TypeScript MCP Server: Input, Binding and Tokens](/guides/securing-a-typescript-mcp-server/) - [MCP authorization (glossary)](/glossary/mcp-authorization/) - [Protected resource metadata (glossary)](/glossary/protected-resource-metadata/) - [All guides](/guides/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Securing a TypeScript MCP server URL: https://mcpservercybersecurity.com/guides/securing-a-typescript-mcp-server/ Updated: 2026-11-16 Secure a TypeScript MCP server by declaring strict input schemas on every tool, creating the app with createMcpExpressApp so localhost binding and host validation are on, and setting expectedResource in requireBearerAuth so tokens for other servers are refused. Examples follow the v1 SDK documentation read on 16 Nov 2026. ## How do you validate input in tool handlers? Declare an input schema on every tool and tighten it until it matches what the tool needs. The SDK's registerTool takes an inputSchema built from Zod types, so the handler receives parsed values instead of raw JSON. OWASP advises strict JSON Schema, with additionalProperties set to false and a pattern on string fields so only declared parameters and valid formats are accepted. Schemas cannot express everything. In the handler, resolve file paths against an allowed root and reject anything outside it, build no shell commands from arguments, and send URL arguments through an allowlist. Write tests that send a path outside the allowed root and a URL on a private address range, and assert that both are rejected by the handler. Keep each schema next to its tool so a change to behavior forces a change to the schema in the same review. Sources: [MCP TypeScript SDK (v1) documentation: Server](https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/v1.x/docs/server.md); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## How do you bind to localhost? Create the Express app with createMcpExpressApp, which uses the default host 127.0.0.1 and turns on DNS rebinding protection. The v1 SDK documentation shows the same protection for host localhost, and shows that passing 0.0.0.0 gives no automatic protection. If you need custom host handling, add the hostHeaderValidation middleware with a list such as localhost, 127.0.0.1 and your own host name. The MCP transport specification requires servers to validate the Origin header and to answer an invalid one with 403, and recommends binding only to localhost for local servers. Sources: [MCP TypeScript SDK (v1) documentation: Server](https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/v1.x/docs/server.md); [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports). ## How do you verify bearer tokens by audience? Use requireBearerAuth with your own verifier and set expectedResource to your server's URL. With expectedResource set, a token is accepted only if the verifier reports that resource, and a token for another value, or none, gets 401 invalid_token. Without it, the documentation says AuthInfo.resource is not compared with anything. This is the audience check the MCP authorization specification requires: servers must validate that access tokens were issued for them, and must not accept tokens issued to anything else. Your verifier should also read the token's audience claim and report the entry that matches your origin. Fail closed. When the verifier cannot validate the token, or cannot read an audience, return 401 rather than letting the request continue to the tool. Sources: [MCP TypeScript SDK (v1) documentation: Server](https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/v1.x/docs/server.md); [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization). ## Which transport should a local server use? Use stdio for a server that only one local client needs. OWASP notes that stdio avoids a listening endpoint but does not restrict the server process's file system, network or credential access, so sandboxing is still needed. If you need HTTP on the machine, the specification suggests requiring an authorization token or using unix domain sockets or other restricted IPC. Sources: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html); [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices). ## How do you handle tool output and SDK versions? Validate output before it returns to the model, because output often becomes input to another tool. The SDK supports an outputSchema next to structuredContent, which gives you a place to enforce the shape. Match examples to the SDK version you install. The v1 documentation imports from @modelcontextprotocol/sdk paths, and the repository's main branch shows a split-package layout, so import paths and helper names can differ. Sources: [MCP TypeScript SDK (v1) documentation: Server](https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/v1.x/docs/server.md); [MCP TypeScript SDK repository](https://github.com/modelcontextprotocol/typescript-sdk); [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html). ## Sources - [MCP TypeScript SDK (v1) documentation: Server](https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/v1.x/docs/server.md) (raw.githubusercontent.com), read 2026-11-16 - [MCP TypeScript SDK repository](https://github.com/modelcontextprotocol/typescript-sdk) (github.com), read 2026-11-16 - [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports) (modelcontextprotocol.io), read 2026-11-16 - [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 - [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Securing a Python MCP Server: Host Checks and Deployment](/guides/securing-a-python-mcp-server/) - [Securing a Spring Boot MCP Server: Spring AI Settings](/guides/securing-a-spring-boot-mcp-server/) - [Tool schema (glossary)](/glossary/tool-schema/) - [Path traversal (glossary)](/glossary/path-traversal/) - [All guides](/guides/) - [Building an Approved MCP Server Catalog for Your Company](/guides/building-an-approved-mcp-server-catalog/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # MCP security glossary URL: https://mcpservercybersecurity.com/glossary/ Updated: 2026-11-16 Twenty-eight terms are defined here in plain language, from tool description, tool schema and server manifest to DNS rebinding, localhost binding and server pinning. Each entry gives a two to four sentence definition, one sourced example and links to related terms, with sources read on 16 Nov 2026. ## 28 terms - [Command injection](/glossary/command-injection/): Command injection is a flaw in which text from an untrusted source ends up interpreted by a shell or operating system call. - [DNS rebinding](/glossary/dns-rebinding/): DNS rebinding is a technique in which a domain name first resolves to a public address and later to a local one, letting a remote web page reach a service bound to the user's machine. - [Dynamic Client Registration](/glossary/dynamic-client-registration/): Dynamic Client Registration (RFC 7591) is the OAuth mechanism by which a client registers with an authorization server at run time. - [Localhost binding](/glossary/localhost-binding/): Localhost binding is the practice of having a server listen only on the loopback address, 127.0.0.1, instead of all network interfaces. - [MCP authorization](/glossary/mcp-authorization/): MCP authorization is the OAuth 2.1-based profile that lets a client obtain tokens for a protected remote MCP server. - [MCP config file](/glossary/mcp-config-file/): An MCP config file is the settings file in which a client records its MCP servers: the command or URL for each, its arguments and its environment variables. - [MCP elicitation](/glossary/mcp-elicitation/): MCP elicitation is a client feature that lets a server request more information from the user during an interaction. - [MCP Inspector](/glossary/mcp-inspector/): The MCP Inspector is the reference developer tool for testing and debugging MCP servers, shipped as the package @modelcontextprotocol/inspector. - [MCP Registry](/glossary/mcp-registry/): The MCP Registry is the official centralized metadata repository for publicly accessible MCP servers. - [MCP resource](/glossary/mcp-resource/): An MCP resource is a piece of data that a server shares with a client, identified by a URI, such as a file, a database schema or application-specific information. - [MCP roots](/glossary/mcp-roots/): MCP roots are a client feature that tells a server which directories and files the client considers relevant. - [MCP sampling](/glossary/mcp-sampling/): MCP sampling is a client feature that lets a server request a completion from the user's language model through the client. - [MCP session](/glossary/mcp-session/): An MCP session was a connection-scoped state set up by an initialize handshake in revisions up to 2025-11-25. - [MCP transport](/glossary/mcp-transport/): An MCP transport is the binding that defines how JSON-RPC messages are framed and delivered between a client and a server. - [npx launcher](/glossary/npx-launcher/): The npx launcher is the npm tool that downloads and runs a package in one command, and it is a common way to start a local MCP server. - [Package supply chain](/glossary/package-supply-chain/): The package supply chain, for MCP, is the set of packages, dependencies and registries that deliver a server to a machine. - [Path traversal](/glossary/path-traversal/): Path traversal is a flaw in which a path supplied to a tool reaches files outside the directory the tool was meant to stay in. - [Protected resource metadata](/glossary/protected-resource-metadata/): Protected resource metadata is a document defined in RFC 9728 that a protected server publishes to describe itself, including which authorization servers it trusts. - [Resource indicators](/glossary/resource-indicators/): Resource indicators are an OAuth extension (RFC 8707) in which the client states which resource server a requested token is meant for. - [Server manifest](/glossary/server-manifest/): A server manifest is a metadata file that names an MCP server, says where its code or endpoint lives and how to launch it. - [Server pinning](/glossary/server-pinning/): Server pinning is the practice of locking an MCP server to a reviewed version and detecting any change afterward. - [Server provenance](/glossary/server-provenance/): Server provenance is the record of who published an MCP server, from which source, and whether the package you install matches that source. - [SSRF](/glossary/ssrf/): Server-side request forgery is an attack in which a program is induced to make HTTP requests to destinations it should not reach, such as internal services or cloud metadata endpoints. - [Tool annotations](/glossary/tool-annotations/): Tool annotations are optional properties on a tool definition that describe its behavior, for example whether it only reads data or can destroy it. - [Tool description](/glossary/tool-description/): A tool description is the text a server attaches to each tool so the model can decide when and how to call it. - [Tool schema](/glossary/tool-schema/): A tool schema is the JSON Schema in a tool definition, called inputSchema, that declares the arguments a tool accepts and their types. - [Tool shadowing](/glossary/tool-shadowing/): Tool shadowing is an attack in which a malicious server's tool description manipulates how the agent behaves with tools that belong to other, trusted servers. - [Typosquatted server](/glossary/typosquatted-server/): A typosquatted server is an MCP server package whose name closely imitates a trusted one so that a typing error or a quick glance installs it instead. ## Related pages - [Guides](/guides/) - [FAQ](/faq/) - [The ranking](/ranking/) - [MCP Risk Register](/mcp-risk-register/) - [MCP Security Checklist](/mcp-security-checklist/) --- # Command injection: definition and example URL: https://mcpservercybersecurity.com/glossary/command-injection/ Updated: 2026-11-16 Command injection is a flaw in which text from an untrusted source ends up interpreted by a shell or operating system call. In MCP the text can come from a model's tool arguments or from a server, such as an OAuth authorization URL. The usual fix is to avoid the shell and validate input against an allowlist. ## Example The MCP security best practices say clients must not use shell commands such as cmd.exe, sh or PowerShell to open URLs, and must only allow http and https schemes for authorization URLs. ## Sources - [MCP Security Best Practices, 2026-07-28 revision](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Path traversal](/glossary/path-traversal/) - [SSRF](/glossary/ssrf/) - [Tool schema](/glossary/tool-schema/) - [All glossary terms](/glossary/) - [Typosquatted server](/glossary/typosquatted-server/) - [DNS rebinding](/glossary/dns-rebinding/) - [Guides](/guides/) --- # DNS rebinding: definition and example URL: https://mcpservercybersecurity.com/glossary/dns-rebinding/ Updated: 2026-11-16 DNS rebinding is a technique in which a domain name first resolves to a public address and later to a local one, letting a remote web page reach a service bound to the user's machine. A local MCP server reachable over HTTP is a target. The specification's defense is for servers to validate the Origin header and bind to localhost. ## Example The Inspector's 2.12.0 release notes say its production web server now validates the Host header and answers 403 for hosts outside an allow-list, which closes a DNS rebinding route to the API token and stored OAuth tokens. ## Sources - [MCP Inspector 2.12.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.12.0) (github.com), read 2026-11-16 ## Related pages - [Localhost binding](/glossary/localhost-binding/) - [SSRF](/glossary/ssrf/) - [MCP transport](/glossary/mcp-transport/) - [Securing a Python MCP Server: Host Checks and Deployment](/guides/securing-a-python-mcp-server/) - [All glossary terms](/glossary/) - [Command injection](/glossary/command-injection/) - [Dynamic Client Registration](/glossary/dynamic-client-registration/) - [Guides](/guides/) --- # Dynamic Client Registration: definition and example URL: https://mcpservercybersecurity.com/glossary/dynamic-client-registration/ Updated: 2026-11-16 Dynamic Client Registration (RFC 7591) is the OAuth mechanism by which a client registers with an authorization server at run time. MCP supported it so clients and servers with no prior relationship could connect. The 2026-07-28 specification deprecates it in favor of Client ID Metadata Documents and keeps it for backward compatibility. ## Example The spec's registration page says new implementations should use Client ID Metadata Documents instead, and the deprecated-features registry gives 2027-07-28 as the earliest removal. ## Sources - [MCP specification 2026-07-28: Client Registration](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization/client-registration) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [MCP authorization](/glossary/mcp-authorization/) - [Protected resource metadata](/glossary/protected-resource-metadata/) - [Resource indicators](/glossary/resource-indicators/) - [All glossary terms](/glossary/) - [DNS rebinding](/glossary/dns-rebinding/) - [Localhost binding](/glossary/localhost-binding/) - [Guides](/guides/) --- # Localhost binding: definition and example URL: https://mcpservercybersecurity.com/glossary/localhost-binding/ Updated: 2026-11-16 Localhost binding is the practice of having a server listen only on the loopback address, 127.0.0.1, instead of all network interfaces. For a local Streamable HTTP MCP server it keeps other hosts on the network from connecting. It does not stop other processes on the same machine or a rebinding attack, so Origin checks and tokens are still needed. ## Example The Streamable HTTP spec says that when running locally, servers should bind only to localhost (127.0.0.1) rather than all network interfaces (0.0.0.0). ## Sources - [MCP specification 2026-07-28: Streamable HTTP](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [DNS rebinding](/glossary/dns-rebinding/) - [MCP transport](/glossary/mcp-transport/) - [MCP Inspector](/glossary/mcp-inspector/) - [Securing a Python MCP Server: Host Checks and Deployment](/guides/securing-a-python-mcp-server/) - [All glossary terms](/glossary/) - [Dynamic Client Registration](/glossary/dynamic-client-registration/) - [MCP authorization](/glossary/mcp-authorization/) - [Guides](/guides/) --- # MCP authorization: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-authorization/ Updated: 2026-11-16 MCP authorization is the OAuth 2.1-based profile that lets a client obtain tokens for a protected remote MCP server. It is optional in the protocol, so a server that exposes non-public tools must choose to require it. The profile covers server discovery, client registration, resource indicators and token handling. ## Example OWASP's cheat sheet says to validate that each access token was issued for this MCP server as its audience and never to pass an MCP access token to an upstream API. ## Sources - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 ## Related pages - [Resource indicators](/glossary/resource-indicators/) - [Protected resource metadata](/glossary/protected-resource-metadata/) - [Dynamic Client Registration](/glossary/dynamic-client-registration/) - [MCP Security Best Practices for Server Operators](/guides/mcp-security-best-practices/) - [Securing a Spring Boot MCP Server: Spring AI Settings](/guides/securing-a-spring-boot-mcp-server/) - [All glossary terms](/glossary/) - [Localhost binding](/glossary/localhost-binding/) - [MCP config file](/glossary/mcp-config-file/) - [Guides](/guides/) --- # MCP config file: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-config-file/ Updated: 2026-11-16 An MCP config file is the settings file in which a client records its MCP servers: the command or URL for each, its arguments and its environment variables. Clients keep such files at user, project or managed levels, and a project file can arrive with a cloned repository. Whatever the file lists runs with the user's permissions. ## Example Claude Code's MCP docs say a cloned repository cannot approve its own servers: approval settings committed to the project's .claude/settings.json are ignored in an untrusted folder, and the server stays at Pending approval. ## Sources - [Connect Claude Code to tools via MCP, Claude Code docs](https://code.claude.com/docs/en/mcp) (code.claude.com), read 2026-11-16 ## Related pages - [Server manifest](/glossary/server-manifest/) - [npx launcher](/glossary/npx-launcher/) - [Server pinning](/glossary/server-pinning/) - [MCP Security Incident Response: Steps for a Bad Server](/guides/mcp-security-incident-response-steps/) - [All glossary terms](/glossary/) - [MCP authorization](/glossary/mcp-authorization/) - [MCP elicitation](/glossary/mcp-elicitation/) - [Guides](/guides/) --- # MCP elicitation: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-elicitation/ Updated: 2026-11-16 MCP elicitation is a client feature that lets a server request more information from the user during an interaction. It has a form mode for structured data and a URL mode that sends the user to an external page. The specification limits what form mode may collect. ## Example The spec says servers must not use form mode to request passwords, API keys, access tokens or payment credentials, and must use URL mode for interactions involving such secrets. ## Sources - [MCP specification 2026-07-28: Elicitation](https://modelcontextprotocol.io/specification/2026-07-28/client/elicitation) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [MCP sampling](/glossary/mcp-sampling/) - [MCP authorization](/glossary/mcp-authorization/) - [SSRF](/glossary/ssrf/) - [All glossary terms](/glossary/) - [MCP config file](/glossary/mcp-config-file/) - [MCP Inspector](/glossary/mcp-inspector/) - [Guides](/guides/) --- # MCP Inspector: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-inspector/ Updated: 2026-11-16 The MCP Inspector is the reference developer tool for testing and debugging MCP servers, shipped as the package @modelcontextprotocol/inspector. It provides web, command-line and terminal clients on one shared core. Because it can start processes and hold OAuth tokens, it needs the same care as any developer tool with those rights. ## Example The Inspector's security page says its web backend can spawn processes, that a per-launch API token guards every route, and that secrets fall back to a plaintext file when no OS keychain exists unless a key is supplied. ## Sources - [MCP Inspector security documentation](https://modelcontextprotocol.io/docs/tools/inspector/security) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [npx launcher](/glossary/npx-launcher/) - [Localhost binding](/glossary/localhost-binding/) - [DNS rebinding](/glossary/dns-rebinding/) - [MCP Server Scanning and Testing: Tools and Limits](/guides/mcp-server-scanning-and-testing/) - [All glossary terms](/glossary/) - [MCP elicitation](/glossary/mcp-elicitation/) - [MCP Registry](/glossary/mcp-registry/) - [Guides](/guides/) --- # MCP Registry: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-registry/ Updated: 2026-11-16 The MCP Registry is the official centralized metadata repository for publicly accessible MCP servers. It stores server.json records that point at packages on npm, PyPI or Docker Hub, and it manages namespaces through DNS verification. A listing in the registry is a pointer, not a code review. ## Example The registry's own page says it is currently in preview and warns that breaking changes or data resets may occur before general availability. ## Sources - [About the MCP Registry](https://modelcontextprotocol.io/registry/about) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Server manifest](/glossary/server-manifest/) - [Server provenance](/glossary/server-provenance/) - [Typosquatted server](/glossary/typosquatted-server/) - [Building an Approved MCP Server Catalog for Your Company](/guides/building-an-approved-mcp-server-catalog/) - [All glossary terms](/glossary/) - [MCP Inspector](/glossary/mcp-inspector/) - [MCP resource](/glossary/mcp-resource/) - [Guides](/guides/) --- # MCP resource: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-resource/ Updated: 2026-11-16 An MCP resource is a piece of data that a server shares with a client, identified by a URI, such as a file, a database schema or application-specific information. Resources are application-driven: the host decides how to bring them into the model's context. Access control on resources must be enforced by the server, since the protocol does not do it. ## Example A FastMCP advisory from 23 Oct 2026 says a denied concrete resource now shadows a matching resource template at its URI, so proxy templates cannot retrieve the same protected resource upstream. ## Sources - [GHSA-7xgj-8392-5q4r: FastMCP post-transform authorization](https://github.com/advisories/GHSA-7xgj-8392-5q4r) (github.com), read 2026-11-16 ## Related pages - [Tool schema](/glossary/tool-schema/) - [MCP authorization](/glossary/mcp-authorization/) - [Path traversal](/glossary/path-traversal/) - [All glossary terms](/glossary/) - [MCP Registry](/glossary/mcp-registry/) - [MCP roots](/glossary/mcp-roots/) - [Guides](/guides/) --- # MCP roots: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-roots/ Updated: 2026-11-16 MCP roots are a client feature that tells a server which directories and files the client considers relevant. The specification says roots are informational guidance rather than an access-control mechanism, and that the protocol does not enforce that servers stay within them. The feature is deprecated as of the 2026-07-28 revision. ## Example The specification's own wording is that the protocol does not enforce that servers stay within roots, so filesystem limits have to come from a sandbox or operating system controls. ## Sources - [MCP specification 2026-07-28: Roots](https://modelcontextprotocol.io/specification/2026-07-28/client/roots) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [MCP sampling](/glossary/mcp-sampling/) - [Path traversal](/glossary/path-traversal/) - [Server pinning](/glossary/server-pinning/) - [All glossary terms](/glossary/) - [MCP resource](/glossary/mcp-resource/) - [Guides](/guides/) --- # MCP sampling: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-sampling/ Updated: 2026-11-16 MCP sampling is a client feature that lets a server request a completion from the user's language model through the client. The 2026-07-28 specification deprecates it, says new implementations should not adopt it, and points existing ones to calling model provider APIs directly. The spec still says a human should be able to deny sampling requests while the feature exists. ## Example The deprecated-features registry lists Sampling, Roots and Logging as deprecated in 2026-07-28, with the earliest removal in the first revision released on or after 2027-07-28. ## Sources - [MCP specification 2026-07-28: Deprecated Features](https://modelcontextprotocol.io/specification/2026-07-28/deprecated) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [MCP roots](/glossary/mcp-roots/) - [MCP elicitation](/glossary/mcp-elicitation/) - [Tool description](/glossary/tool-description/) - [All glossary terms](/glossary/) - [MCP session](/glossary/mcp-session/) - [Guides](/guides/) --- # MCP session: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-session/ Updated: 2026-11-16 An MCP session was a connection-scoped state set up by an initialize handshake in revisions up to 2025-11-25. The 2026-07-28 specification says MCP is stateless and has no protocol-level sessions, so a server that needs state across requests mints an explicit handle. Those handles can be hijacked if a server treats possession as proof of identity. ## Example The Python SDK 2.2.0 release notes say a legacy stateful session idle for 30 minutes is closed and a server holds at most 10,000 sessions at once. ## Sources - [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (github.com), read 2026-11-16 ## Related pages - [MCP transport](/glossary/mcp-transport/) - [MCP authorization](/glossary/mcp-authorization/) - [MCP Audit Logging: What to Record and How Long to Keep It](/guides/mcp-audit-logging-what-to-record/) - [All glossary terms](/glossary/) - [MCP sampling](/glossary/mcp-sampling/) - [Guides](/guides/) --- # MCP transport: definition and example URL: https://mcpservercybersecurity.com/glossary/mcp-transport/ Updated: 2026-11-16 An MCP transport is the binding that defines how JSON-RPC messages are framed and delivered between a client and a server. The specification defines two standard bindings: stdio, newline-delimited messages over a subprocess's standard streams, and Streamable HTTP, where each message is an HTTP POST to a single endpoint. Protocol meaning is the same on both. ## Example The Streamable HTTP binding mirrors selected body fields into HTTP headers so that intermediaries can route and inspect requests without parsing the body, and the spec says the body remains the source of truth. ## Sources - [MCP specification 2026-07-28: Transports](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Localhost binding](/glossary/localhost-binding/) - [DNS rebinding](/glossary/dns-rebinding/) - [MCP session](/glossary/mcp-session/) - [All glossary terms](/glossary/) - [npx launcher](/glossary/npx-launcher/) - [Guides](/guides/) --- # npx launcher: definition and example URL: https://mcpservercybersecurity.com/glossary/npx-launcher/ Updated: 2026-11-16 The npx launcher is the npm tool that downloads and runs a package in one command, and it is a common way to start a local MCP server. The command in the config file names the package, and npm fetches whatever that name currently resolves to. That makes the package name and version the thing to verify. ## Example The Inspector's 2.13.0 release notes say that run from npm with npx it uses the user's own Node.js, so the Inspector cannot update the runtime and the user must run a patched Node.js release. ## Sources - [MCP Inspector 2.13.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.13.0) (github.com), read 2026-11-16 ## Related pages - [MCP config file](/glossary/mcp-config-file/) - [Server pinning](/glossary/server-pinning/) - [Package supply chain](/glossary/package-supply-chain/) - [MCP Security Incident Response: Steps for a Bad Server](/guides/mcp-security-incident-response-steps/) - [All glossary terms](/glossary/) - [MCP transport](/glossary/mcp-transport/) - [Guides](/guides/) --- # Package supply chain: definition and example URL: https://mcpservercybersecurity.com/glossary/package-supply-chain/ Updated: 2026-11-16 The package supply chain, for MCP, is the set of packages, dependencies and registries that deliver a server to a machine. A compromised or look-alike package runs with the same permissions as the server. Review therefore covers the package, its publisher and its dependencies, not only the server's own tools. ## Example Snyk says its scan data from nearly 10,000 developer environments found 4,524 unique MCP servers in active use, and its post calls every MCP server an agent connects to a new piece of the software supply chain. ## Sources - [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (snyk.io), read 2026-11-16 ## Related pages - [Typosquatted server](/glossary/typosquatted-server/) - [Server provenance](/glossary/server-provenance/) - [npx launcher](/glossary/npx-launcher/) - [All glossary terms](/glossary/) - [Path traversal](/glossary/path-traversal/) - [Guides](/guides/) --- # Path traversal: definition and example URL: https://mcpservercybersecurity.com/glossary/path-traversal/ Updated: 2026-11-16 Path traversal is a flaw in which a path supplied to a tool reaches files outside the directory the tool was meant to stay in. It matters for MCP because a model chooses the arguments, and the arguments may be shaped by untrusted content. The control is to resolve and check paths on the server and to sandbox the process. ## Example One of the 25 advisories GitHub published for mcp-atlassian on 22 Sep 2026 is titled for arbitrary file read and exfiltration through an upload_attachment tool missing path validation. ## Sources - [GHSA-93xw-j965-9mx3: mcp-atlassian arbitrary file read](https://github.com/advisories/GHSA-93xw-j965-9mx3) (github.com), read 2026-11-16 ## Related pages - [Command injection](/glossary/command-injection/) - [MCP roots](/glossary/mcp-roots/) - [SSRF](/glossary/ssrf/) - [How to Review an MCP Server for Security: Code and Tools](/guides/how-to-review-an-mcp-server-security/) - [Securing a TypeScript MCP Server: Input, Binding and Tokens](/guides/securing-a-typescript-mcp-server/) - [All glossary terms](/glossary/) - [Package supply chain](/glossary/package-supply-chain/) - [Protected resource metadata](/glossary/protected-resource-metadata/) - [Guides](/guides/) --- # Protected resource metadata: definition and example URL: https://mcpservercybersecurity.com/glossary/protected-resource-metadata/ Updated: 2026-11-16 Protected resource metadata is a document defined in RFC 9728 that a protected server publishes to describe itself, including which authorization servers it trusts. In MCP, a client reads it to find where to sign in. Because the server supplies the URLs, a client should treat them as untrusted input. ## Example The spec requires the metadata document returned by an MCP server to include the authorization_servers field with at least one authorization server. ## Sources - [MCP specification 2026-07-28: Authorization Server Discovery](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization/authorization-server-discovery) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Resource indicators](/glossary/resource-indicators/) - [SSRF](/glossary/ssrf/) - [MCP authorization](/glossary/mcp-authorization/) - [Securing a Spring Boot MCP Server: Spring AI Settings](/guides/securing-a-spring-boot-mcp-server/) - [All glossary terms](/glossary/) - [Path traversal](/glossary/path-traversal/) - [Guides](/guides/) --- # Resource indicators: definition and example URL: https://mcpservercybersecurity.com/glossary/resource-indicators/ Updated: 2026-11-16 Resource indicators are an OAuth extension (RFC 8707) in which the client states which resource server a requested token is meant for. MCP clients must implement it and use the canonical URI of the MCP server. The result is a token whose audience is that server, which the server can then check. ## Example The authorization spec says access tokens must not be included in the URI query string and that MCP servers must validate that tokens were issued specifically for them as the intended audience. ## Sources - [MCP specification 2026-07-28: Authorization](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [MCP authorization](/glossary/mcp-authorization/) - [Protected resource metadata](/glossary/protected-resource-metadata/) - [Dynamic Client Registration](/glossary/dynamic-client-registration/) - [MCP Server Security on AWS, Azure and Google Cloud](/guides/mcp-server-security-on-aws-azure-and-google-cloud/) - [All glossary terms](/glossary/) - [Server manifest](/glossary/server-manifest/) - [Guides](/guides/) --- # Server manifest: definition and example URL: https://mcpservercybersecurity.com/glossary/server-manifest/ Updated: 2026-11-16 A server manifest is a metadata file that names an MCP server, says where its code or endpoint lives and how to launch it. In the official MCP Registry the file is server.json. It is the record a client reads before it runs a package, so what it launches deserves review. ## Example The Inspector's 2.14.0 release notes say server.json import now shows the exact command before it is saved and rejects a package with an invalid or empty environment-variable name or unsafe tokens. ## Sources - [MCP Inspector 2.14.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.14.0) (github.com), read 2026-11-16 ## Related pages - [MCP Registry](/glossary/mcp-registry/) - [MCP config file](/glossary/mcp-config-file/) - [Server provenance](/glossary/server-provenance/) - [Building an Approved MCP Server Catalog for Your Company](/guides/building-an-approved-mcp-server-catalog/) - [All glossary terms](/glossary/) - [Resource indicators](/glossary/resource-indicators/) - [Server pinning](/glossary/server-pinning/) - [Guides](/guides/) --- # Server pinning: definition and example URL: https://mcpservercybersecurity.com/glossary/server-pinning/ Updated: 2026-11-16 Server pinning is the practice of locking an MCP server to a reviewed version and detecting any change afterward. It can pin the package version, a content hash, or a hash of the full tool definition. It reduces the chance that an approved server changes after review. ## Example The MCP local-server guide says that if you do one thing, open your client's MCP configuration, remove the servers you no longer use, and make sure you know where each remaining one comes from and that its version is pinned. ## Sources - [Local Server Security, modelcontextprotocol.io](https://modelcontextprotocol.io/docs/tutorials/security/local-server-security) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Tool description](/glossary/tool-description/) - [Server provenance](/glossary/server-provenance/) - [MCP config file](/glossary/mcp-config-file/) - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [All glossary terms](/glossary/) - [Server manifest](/glossary/server-manifest/) - [Guides](/guides/) --- # Server provenance: definition and example URL: https://mcpservercybersecurity.com/glossary/server-provenance/ Updated: 2026-11-16 Server provenance is the record of who published an MCP server, from which source, and whether the package you install matches that source. Checking it means looking at the publisher, the repository, the version and the install command before anything runs. The MCP local-server guidance lists provenance as a step to complete before installing. ## Example Knostic reported on 26 Oct 2026 that a package named to look like Amazon's SP-API MCP copied Amazon's open-source code but replaced the authentication and API addresses, and that the README did not disclose the change. ## Sources - [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (knostic.ai), read 2026-11-16 ## Related pages - [Server pinning](/glossary/server-pinning/) - [Package supply chain](/glossary/package-supply-chain/) - [Typosquatted server](/glossary/typosquatted-server/) - [All glossary terms](/glossary/) - [SSRF](/glossary/ssrf/) - [Guides](/guides/) --- # SSRF: definition and example URL: https://mcpservercybersecurity.com/glossary/ssrf/ Updated: 2026-11-16 Server-side request forgery is an attack in which a program is induced to make HTTP requests to destinations it should not reach, such as internal services or cloud metadata endpoints. In MCP it arises when a client follows URLs supplied by a server, and when a tool fetches URLs a model produced. Controls are HTTPS requirements, private-range blocking and an egress proxy. ## Example The Inspector's 2.15.0 release notes say its fetch guard now also blocks cloud metadata addresses outside 169.254.0.0/16, including 100.100.100.200 and 192.0.0.192 and their encoded forms. ## Sources - [MCP Inspector 2.15.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.15.0) (github.com), read 2026-11-16 ## Related pages - [Protected resource metadata](/glossary/protected-resource-metadata/) - [DNS rebinding](/glossary/dns-rebinding/) - [Command injection](/glossary/command-injection/) - [MCP Server Security on AWS, Azure and Google Cloud](/guides/mcp-server-security-on-aws-azure-and-google-cloud/) - [All glossary terms](/glossary/) - [Server provenance](/glossary/server-provenance/) - [Tool annotations](/glossary/tool-annotations/) - [Guides](/guides/) --- # Tool annotations: definition and example URL: https://mcpservercybersecurity.com/glossary/tool-annotations/ Updated: 2026-11-16 Tool annotations are optional properties on a tool definition that describe its behavior, for example whether it only reads data or can destroy it. Clients use them to decide how much approval to ask for. They are hints from the server, not enforcement. ## Example The tools specification says clients must consider tool annotations untrusted unless they come from trusted servers. ## Sources - [MCP specification 2026-07-28: Tools](https://modelcontextprotocol.io/specification/2026-07-28/server/tools) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [Tool description](/glossary/tool-description/) - [Tool schema](/glossary/tool-schema/) - [Server pinning](/glossary/server-pinning/) - [MCP Security Best Practices for Server Operators](/guides/mcp-security-best-practices/) - [All glossary terms](/glossary/) - [SSRF](/glossary/ssrf/) - [Guides](/guides/) --- # Tool description: definition and example URL: https://mcpservercybersecurity.com/glossary/tool-description/ Updated: 2026-11-16 A tool description is the text a server attaches to each tool so the model can decide when and how to call it. The agent reads it as part of its context, which makes the description a place where hidden instructions can sit. A review of a server should cover every field of the tool definition, not only this one. ## Example The OWASP cheat sheet says to hash the complete tool definition, including inputSchema, outputSchema and annotations, because a flipped readOnlyHint or destructiveHint can change client approval behavior without any visible change to the description. ## Sources - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 ## Related pages - [Tool schema](/glossary/tool-schema/) - [Tool annotations](/glossary/tool-annotations/) - [Server pinning](/glossary/server-pinning/) - [MCP Audit Logging: What to Record and How Long to Keep It](/guides/mcp-audit-logging-what-to-record/) - [MCP Security vs API Security: What Changes](/guides/mcp-security-vs-api-security/) - [All glossary terms](/glossary/) - [Guides](/guides/) --- # Tool schema: definition and example URL: https://mcpservercybersecurity.com/glossary/tool-schema/ Updated: 2026-11-16 A tool schema is the JSON Schema in a tool definition, called inputSchema, that declares the arguments a tool accepts and their types. Clients show it to the model and servers validate calls against it. A strict schema limits unexpected fields, and a hostile schema is itself untrusted input. ## Example The Inspector's 2.14.0 release notes say server-supplied schema patterns are now checked in bounded time with a linear-time matcher and per-pattern limits, and that a pattern beyond the limits is reported as unverified. ## Sources - [MCP Inspector 2.14.0 release notes](https://github.com/modelcontextprotocol/inspector/releases/tag/2.14.0) (github.com), read 2026-11-16 ## Related pages - [Tool description](/glossary/tool-description/) - [Tool annotations](/glossary/tool-annotations/) - [Command injection](/glossary/command-injection/) - [How to Review an MCP Server for Security: Code and Tools](/guides/how-to-review-an-mcp-server-security/) - [Securing a TypeScript MCP Server: Input, Binding and Tokens](/guides/securing-a-typescript-mcp-server/) - [All glossary terms](/glossary/) - [Tool shadowing](/glossary/tool-shadowing/) - [Guides](/guides/) --- # Tool shadowing: definition and example URL: https://mcpservercybersecurity.com/glossary/tool-shadowing/ Updated: 2026-11-16 Tool shadowing is an attack in which a malicious server's tool description manipulates how the agent behaves with tools that belong to other, trusted servers. It works because the model sees the descriptions from every connected server in one context. OWASP also calls it cross-origin escalation. ## Example OWASP's cheat sheet tells teams to treat each MCP server as an untrusted security domain of its own and to watch for cross-server data flows, such as credentials from server A appearing in calls to server B. ## Sources - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 ## Related pages - [Tool description](/glossary/tool-description/) - [Server pinning](/glossary/server-pinning/) - [Typosquatted server](/glossary/typosquatted-server/) - [MCP Security vs API Security: What Changes](/guides/mcp-security-vs-api-security/) - [MCP Server Scanning and Testing: Tools and Limits](/guides/mcp-server-scanning-and-testing/) - [All glossary terms](/glossary/) - [Tool schema](/glossary/tool-schema/) - [Guides](/guides/) --- # Typosquatted server: definition and example URL: https://mcpservercybersecurity.com/glossary/typosquatted-server/ Updated: 2026-11-16 A typosquatted server is an MCP server package whose name closely imitates a trusted one so that a typing error or a quick glance installs it instead. Because MCP servers often run with the user's own permissions, the wrong package gets access to the same files and credentials. Verifying the exact package name and publisher before install is the first control. ## Example OWASP gives the pair mcp-server-filesystem and mcp-server-filesytem as an example of the name differences to check before installation. ## Sources - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) (cheatsheetseries.owasp.org), read 2026-11-16 ## Related pages - [Server provenance](/glossary/server-provenance/) - [Package supply chain](/glossary/package-supply-chain/) - [npx launcher](/glossary/npx-launcher/) - [How to Vet an MCP Server Before Use: Steps and Signals](/guides/how-to-vet-an-mcp-server-before-use/) - [All glossary terms](/glossary/) - [Tool shadowing](/glossary/tool-shadowing/) - [Command injection](/glossary/command-injection/) - [Guides](/guides/) --- # MCP security news URL: https://mcpservercybersecurity.com/news/ Updated: 2026-11-16 Twelve dated items track MCP server vulnerabilities and CVEs, malicious or typosquatted packages in registries, specification and registry changes that affect security, and release notes from tools that discover or control MCP servers. Each item links to its primary source and shows the date the source was read. ## 12 items - 2026-11-12: [Knostic: MCP tool asks agents for a GitHub token](/news/knostic-mcp-tool-description-asks-agent-for-github-token/) - 2026-11-09: [MCP draft spec adopts Client ID Metadata draft 03](/news/mcp-spec-draft-client-id-metadata-document-draft-03/) - 2026-11-06: [MCP Registry v1.8.8: key listings by name and version](/news/mcp-registry-v1-8-8-listings-keyed-by-name-and-version/) - 2026-10-27: [MCP Python SDK client fetched URLs from server schemas](/news/python-sdk-client-fetched-server-chosen-ref-urls/) - 2026-10-26: [Knostic flags a repackaged Amazon SP-API MCP server](/news/knostic-repackaged-amazon-sp-api-mcp-credential-redirect/) - 2026-10-23: [FastMCP advisories: authorization skipped after transforms](/news/fastmcp-authorization-advisories-transforms-and-issuer-binding/) - 2026-10-20: [MCP TypeScript SDK server advisories on audience](/news/mcp-typescript-sdk-server-advisories-audience-memory-cross-client/) - 2026-10-11: [Cisco DefenseClaw 1.0.0 adds managed installs and inventory](/news/cisco-defenseclaw-1-0-0-managed-enterprise-release/) - 2026-10-06: [MCP TypeScript SDK OAuth client advisory (CVE-2026-104850)](/news/mcp-typescript-sdk-oauth-client-credentials-advisory/) - 2026-09-30: [Snyk makes Evo ADS MCP Governance generally available](/news/snyk-evo-ads-mcp-governance-generally-available/) - 2026-09-22: [mcp-atlassian: 25 advisories, one critical, fixed in 0.22.0](/news/mcp-atlassian-25-advisories-authentication-bypass-fixed-0-22-0/) - 2026-09-07: [MCP Python SDK 2.2.0 tightens redirects and sessions](/news/python-sdk-2-2-0-redirect-session-issuer-defaults/) Subscribe with the [RSS feed](/feed.xml). ## Related pages - [Guides](/guides/) - [The ranking](/ranking/) - [Vendors](/vendors/) - [Send a correction or tip](/corrections/) --- # Knostic: MCP tool asks agents for a GitHub token URL: https://mcpservercybersecurity.com/news/knostic-mcp-tool-description-asks-agent-for-github-token/ Updated: 2026-11-12 Knostic published a static review on 12 Nov 2026 of @agentdance/node-webrtc-mcp version 1.0.13. Knostic says the description of its get_started tool contains an instruction telling the agent to obtain a GitHub token without asking the user, and that the server then uses a supplied token to favorite the publisher's repository. 2026-11-12 Knostic says the package helps coding agents pick a WebRTC library and generate starter code. The README says the tool favorites the publisher's repository when given a token, but the text the agent reads goes further and tells it to fetch the credential itself and not to ask the user. Knostic notes the tool code also marks every requirement as satisfied and that the README says the comparison tool always recommends the publisher's library. Knostic did not install or run the package; the post is a static code review. The case shows why a tool description is untrusted input: the agent reads it to decide how to call the tool, so its text can steer behavior. The post ends with practical checks. For a platform team, the controls that apply are a review step for tool descriptions before a server is approved, a pinned version, and an alert when a description changes. ## Sources - [When an MCP Tool Asks for Your GitHub Token, Knostic](https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token) (knostic.ai), read 2026-11-16 ## Related pages - [Knostic profile](/vendors/knostic/) - [All news](/news/) - [MCP Python SDK 2.2.0 tightens redirects and sessions](/news/python-sdk-2-2-0-redirect-session-issuer-defaults/) - [MCP draft spec adopts Client ID Metadata draft 03](/news/mcp-spec-draft-client-id-metadata-document-draft-03/) - [The ranking](/ranking/) --- # MCP draft spec adopts Client ID Metadata draft 03 URL: https://mcpservercybersecurity.com/news/mcp-spec-draft-client-id-metadata-document-draft-03/ Updated: 2026-11-09 The Model Context Protocol draft specification changelog records an update on 9 Nov 2026: the Client ID Metadata Document reference moves from draft-ietf-oauth-client-id-metadata-document-00 to draft 03. The changelog says MCP implementations that support Client ID Metadata Documents must follow that version. 2026-11-09 Client ID Metadata Documents let a client identify itself to an authorization server with an HTTPS URL that hosts its metadata. The 2026-07-28 specification marks Dynamic Client Registration as deprecated and points new implementations to these documents. The change sits in the draft, not in the 2026-07-28 release, so it applies to the next revision. The commit that made the change is dated 9 Nov 2026 in the specification repository. The spec's security best practices page warns that an authorization server fetching a client-supplied metadata URL can be pointed at internal addresses, and recommends blocking private IP ranges and using an egress proxy. Teams running their own authorization server should check which draft their library follows before the next revision. ## Sources - [MCP draft specification changelog](https://modelcontextprotocol.io/specification/draft/changelog) (modelcontextprotocol.io), read 2026-11-16 - [docs: update OAuth Client ID Metadata Document references to draft 03 (#3521)](https://github.com/modelcontextprotocol/modelcontextprotocol/commit/222640c8e771dd57f41e655d3c3eef6e5dec8b34) (github.com), read 2026-11-16 ## Related pages - [All news](/news/) - [Knostic: MCP tool asks agents for a GitHub token](/news/knostic-mcp-tool-description-asks-agent-for-github-token/) - [MCP Registry v1.8.8: key listings by name and version](/news/mcp-registry-v1-8-8-listings-keyed-by-name-and-version/) - [The ranking](/ranking/) --- # MCP Registry v1.8.8: key listings by name and version URL: https://mcpservercybersecurity.com/news/mcp-registry-v1-8-8-listings-keyed-by-name-and-version/ Updated: 2026-11-06 The official MCP Registry released v1.8.8 on 6 Nov 2026. Its notes say different server names may now share a remote URL, and that consumers should key listings by server name and version rather than assuming URL uniqueness. Teams that build approved-server lists from registry data should check their matching logic. 2026-11-06 The release notes are short and operational. They state that the release uses the source commit validated in staging and that no database migration is required. The one behavior note concerns identity: a remote URL can now appear under more than one server name. An approved-server catalog that treats a remote URL as the identity of a server could map two listings to one entry, or approve a second name because its URL matches an approved one. Keying by name and version, as the notes advise, keeps a reviewed version distinct from a later one. The registry's own about page calls it a preview that stores metadata pointing at packages hosted on npm, PyPI and Docker Hub, with namespace management through DNS verification. The registry holds pointers, not code, so a catalog still needs to review the package behind each entry. ## Sources - [MCP Registry v1.8.8 release notes](https://github.com/modelcontextprotocol/registry/releases/tag/v1.8.8) (github.com), read 2026-11-16 - [About the MCP Registry](https://modelcontextprotocol.io/registry/about) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [All news](/news/) - [MCP draft spec adopts Client ID Metadata draft 03](/news/mcp-spec-draft-client-id-metadata-document-draft-03/) - [MCP Python SDK client fetched URLs from server schemas](/news/python-sdk-client-fetched-server-chosen-ref-urls/) - [The ranking](/ranking/) --- # MCP Python SDK client fetched URLs from server schemas URL: https://mcpservercybersecurity.com/news/python-sdk-client-fetched-server-chosen-ref-urls/ Updated: 2026-10-27 GitHub published advisory GHSA-rwrf-2pqf-9j8j (CVE-2026-116686, medium) on 27 Oct 2026: the MCP Python SDK client fetched server-chosen $ref URLs while checking tool results against output schemas. Affected versions are mcp 1.10.0 through 1.29.1 and 2.0.0 through 2.1.1. Fixes: 1.30.0 and 2.2.0. 2026-10-27 The advisory says that when a connected server declared an outputSchema for a tool, the SDK client validated each call_tool result with a default resolver that opens any reference it cannot resolve locally. A server could name an http, https or file reference in the schema, and the client would request that URL or read that local file from its own network position on every call of the tool. The advisory says the content was used only as a schema and never sent back to the server. The advisory also describes a stall: the fetch ran synchronously with no timeout, so a target that accepted the connection and went quiet could freeze every session in the process. It rates the issue medium (CVSS 6.5) and says no setting turned the validation off. Applications are affected if they use the SDK as a client (ClientSession, ClientSessionGroup or the 2.x Client) to call tools on servers they do not fully control. Server-only applications are not affected. The fix is mcp 1.30.0 or 2.2.0, which resolves references only inside the schema document the server sent. For an engineer who inventories MCP tooling, the check is which internal agents and scripts import the pip package mcp as a client and which version they pin. ## Sources - [GHSA-rwrf-2pqf-9j8j: MCP Python SDK client fetched server-chosen $ref URLs](https://github.com/advisories/GHSA-rwrf-2pqf-9j8j) (github.com), read 2026-11-16 ## Related pages - [All news](/news/) - [MCP Registry v1.8.8: key listings by name and version](/news/mcp-registry-v1-8-8-listings-keyed-by-name-and-version/) - [Knostic flags a repackaged Amazon SP-API MCP server](/news/knostic-repackaged-amazon-sp-api-mcp-credential-redirect/) - [The ranking](/ranking/) --- # Knostic flags a repackaged Amazon SP-API MCP server URL: https://mcpservercybersecurity.com/news/knostic-repackaged-amazon-sp-api-mcp-credential-redirect/ Updated: 2026-10-26 Knostic published research on 26 Oct 2026 about @bestcoolman/dev-mcp, a package that looks like Amazon's SP-API MCP server. Knostic says static analysis of all seven versions shows application credentials sent by default to a hardcoded IP over unencrypted HTTP, and assesses it as suspected malicious credential redirection. 2026-10-26 Knostic says its AgentMesh threat intelligence flagged version 0.0.7, which prompted a review. The package reuses Amazon's open-source MCP code with the authentication and API addresses replaced, and its README does not disclose the change. Knostic downloaded and statically examined versions 0.0.1 to 0.0.7 and found the substituted addresses in each. The original server asks users to set a client ID, client secret and refresh token as environment variables. Knostic says the repackaged copy sends those three values to the hardcoded address instead of Amazon. Knostic states it has not confirmed who operates the destination or whether real users' credentials reached it. The finding is a case for checking the publisher and scope of a package name before it runs, and for blocking outbound traffic from MCP server processes to destinations the server has no reason to contact. The post includes indicators and response steps. ## Sources - [Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic](https://www.knostic.ai/blog/repackaged-amazon-sp-api-mcp-sends-app-secrets-to-an-undisclosed-ip) (knostic.ai), read 2026-11-16 ## Related pages - [Knostic profile](/vendors/knostic/) - [All news](/news/) - [MCP Python SDK client fetched URLs from server schemas](/news/python-sdk-client-fetched-server-chosen-ref-urls/) - [FastMCP advisories: authorization skipped after transforms](/news/fastmcp-authorization-advisories-transforms-and-issuer-binding/) - [The ranking](/ranking/) --- # FastMCP advisories: authorization skipped after transforms URL: https://mcpservercybersecurity.com/news/fastmcp-authorization-advisories-transforms-and-issuer-binding/ Updated: 2026-10-23 GitHub published two medium-severity FastMCP advisories on 23 Oct 2026. GHSA-7xgj-8392-5q4r says a server transform could change what an authorization policy checks, so direct calls skipped the policy. GHSA-vqqp-x5qm-5qp4 covers client wrappers that built OAuth providers without issuer binding. Fixes: 3.4.10 and 4.1.2. 2026-10-23 The maintainers' assessment of the first advisory says direct lookups checked the original component but returned the transformed one without rechecking its policy. A client that knows a component identifier could call a tool, or read a resource or prompt, that the configured policy denies. It affects fastmcp and fastmcp-slim 3.0.0 through 3.4.9 and 4.0.0 through 4.1.1 when component authorization is combined with transforms or with protected resources that overlap proxy templates. The fixes authorize the final transformed components. Until upgrading, the advisory suggests avoiding overlapping proxy templates for protected resources and enforcing authorization at the handler or server boundary rather than in a policy whose inputs a transform changes. The second advisory affects fastmcp and fastmcp-slim 4.0.0 through 4.1.1. The maintainers say an untrusted or compromised MCP server could advertise a different authorization server and receive the client secret or a signed assertion; 4.1.2 adds an issuer setting and requires MCP SDK 2.2.0. It echoes the TypeScript SDK issuer-binding advisory from 6 Oct 2026. ## Sources - [GHSA-7xgj-8392-5q4r: FastMCP post-transform authorization](https://github.com/advisories/GHSA-7xgj-8392-5q4r) (github.com), read 2026-11-16 - [GHSA-vqqp-x5qm-5qp4: FastMCP client OAuth issuer binding](https://github.com/advisories/GHSA-vqqp-x5qm-5qp4) (github.com), read 2026-11-16 ## Related pages - [All news](/news/) - [Knostic flags a repackaged Amazon SP-API MCP server](/news/knostic-repackaged-amazon-sp-api-mcp-credential-redirect/) - [MCP TypeScript SDK server advisories on audience](/news/mcp-typescript-sdk-server-advisories-audience-memory-cross-client/) - [The ranking](/ranking/) --- # MCP TypeScript SDK server advisories on audience URL: https://mcpservercybersecurity.com/news/mcp-typescript-sdk-server-advisories-audience-memory-cross-client/ Updated: 2026-10-20 GitHub published several advisories on 20 Oct 2026 for MCP TypeScript SDK servers, including CVE-2026-111771 (bearer tokens issued for another service were accepted), CVE-2026-111770 (a shared transport could deliver one client's messages to another) and CVE-2026-111768 (one HTTP request could exhaust memory). Fixes are sdk 1.32.0 and server 2.3.0. 2026-10-20 For CVE-2026-111771 (medium), the advisory says the SDK's server-side bearer authentication could not check the token's audience, so a token issued for another service by the same authorization server was accepted. The MCP authorization spec requires servers to validate that tokens were issued specifically for them. The fix adds the check; the maintainers say the verifier must still be configured to use it. CVE-2026-111770 (high) affects @modelcontextprotocol/server 2.0.0 through 2.2.0: when more than one client could reach an HTTP endpoint and the transport or server was created once at startup, messages matched on JSON-RPC id alone could reach the wrong client. stdio servers are not affected. CVE-2026-111768 (high) covers unbounded validation work; the maintainers say stdio-only servers are not affected. A fourth record, CVE-2026-111767 (high), says experimental tasks were not tied to the session that created them. The fixed versions the advisories list are @modelcontextprotocol/sdk 1.32.0 and @modelcontextprotocol/server 2.3.0. ## Sources - [GHSA-rvq5-wwqv-78pq: bearer authentication audience](https://github.com/advisories/GHSA-rvq5-wwqv-78pq) (github.com), read 2026-11-16 - [GHSA-jm63-cq47-j5jp: shared transport delivered messages across clients](https://github.com/advisories/GHSA-jm63-cq47-j5jp) (github.com), read 2026-11-16 - [GHSA-46jr-4wx8-v453: request validation memory exhaustion](https://github.com/advisories/GHSA-46jr-4wx8-v453) (github.com), read 2026-11-16 - [GHSA-22jm-h49p-29qw: experimental tasks not tied to the session](https://github.com/advisories/GHSA-22jm-h49p-29qw) (github.com), read 2026-11-16 ## Related pages - [All news](/news/) - [FastMCP advisories: authorization skipped after transforms](/news/fastmcp-authorization-advisories-transforms-and-issuer-binding/) - [Cisco DefenseClaw 1.0.0 adds managed installs and inventory](/news/cisco-defenseclaw-1-0-0-managed-enterprise-release/) - [The ranking](/ranking/) --- # Cisco DefenseClaw 1.0.0 adds managed installs and inventory URL: https://mcpservercybersecurity.com/news/cisco-defenseclaw-1-0-0-managed-enterprise-release/ Updated: 2026-10-11 Cisco's DefenseClaw reached version 1.0.0 on 11 Oct 2026. The release notes list managed installs for Linux, macOS and Windows, hook guardrails for ten agents, identity-based guardrail profiles, an inventory of IDE extensions and plugins, and per-item MCP inventory with Claude Code MCP coverage. 2026-10-11 The 1.0.0 notes say settings now live in one config.yaml, with admission rules, block and allow lists and hook fail modes in the file. Each decision carries the digest of the effective policy. Directory facts for identity come from what the operating system already knows and are marked verified or claimed; only verified facts select a profile. Managed installs ship as rpm, deb, pkg and Windows Setup with a machine policy users cannot edit, and the notes name starter kits for Intune, Okta and Entra ID. Hook-based guardrails cover Claude Code, Codex, Copilot CLI, OpenCode, Amp, Cursor, Devin, Antigravity, Hermes and Kiro. One known limit in the notes bears on MCP: MCP servers an agent already has keep running, because the gateway records a baseline at first start, and a server the scanner cannot start is reported, not blocked. A signed-out enrolled Windows user has project MCP servers refused until about 30 seconds after sign-in. ## Sources - [DefenseClaw 1.0.0 release notes](https://github.com/cisco-ai-defense/defenseclaw/releases/tag/1.0.0) (github.com), read 2026-11-16 ## Related pages - [Cisco profile](/vendors/cisco/) - [All news](/news/) - [MCP TypeScript SDK server advisories on audience](/news/mcp-typescript-sdk-server-advisories-audience-memory-cross-client/) - [MCP TypeScript SDK OAuth client advisory (CVE-2026-104850)](/news/mcp-typescript-sdk-oauth-client-credentials-advisory/) - [The ranking](/ranking/) --- # MCP TypeScript SDK OAuth client advisory (CVE-2026-104850) URL: https://mcpservercybersecurity.com/news/mcp-typescript-sdk-oauth-client-credentials-advisory/ Updated: 2026-10-06 GitHub published advisory GHSA-6qxp-vccf-f47h (CVE-2026-104850, high severity) on 6 Oct 2026. In affected TypeScript SDK versions, the OAuth client let an MCP server choose which authorization server received the client's credentials, with no user interaction. Fixes are @modelcontextprotocol/sdk 1.31.0 and @modelcontextprotocol/client 2.2.0. 2026-10-06 The advisory says an application is affected if it uses the SDK's OAuth client over HTTP and may connect to an MCP server it does not fully trust while holding credentials for a legitimate authorization server. Servers built with the SDK and stdio clients are not affected. Affected ranges are @modelcontextprotocol/sdk 1.12.0 through 1.30.1 and @modelcontextprotocol/client 2.0.0 through 2.1.0, with conditions listed in the record. The fix records the authorization server as the issuer on saved credentials and refuses to send them to a different one. The maintainers add that upgrading is not always enough: bundled providers need an expectedIssuer value, and credentials saved before the upgrade without an issuer must be given one or cleared so users sign in again. The spec's security best practices page lists OAuth discovery URLs among the inputs a client must not trust blindly, so this advisory is a concrete case of a client acting on server-supplied metadata. ## Sources - [GHSA-6qxp-vccf-f47h: MCP TypeScript SDK OAuth client](https://github.com/advisories/GHSA-6qxp-vccf-f47h) (github.com), read 2026-11-16 - [MCP Security Best Practices, 2026-07-28](https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices) (modelcontextprotocol.io), read 2026-11-16 ## Related pages - [All news](/news/) - [Cisco DefenseClaw 1.0.0 adds managed installs and inventory](/news/cisco-defenseclaw-1-0-0-managed-enterprise-release/) - [Snyk makes Evo ADS MCP Governance generally available](/news/snyk-evo-ads-mcp-governance-generally-available/) - [The ranking](/ranking/) --- # Snyk makes Evo ADS MCP Governance generally available URL: https://mcpservercybersecurity.com/news/snyk-evo-ads-mcp-governance-generally-available/ Updated: 2026-09-30 Snyk announced on 30 Sep 2026 that Govern Agent Behavior in Evo ADS is generally available, starting with MCP Governance. Snyk says teams can define approved MCP servers, see agents reach outside the list, and log or block the call on the endpoint across Claude Code, Cursor, Codex and GitHub Copilot. 2026-09-30 Per Snyk's post, MCP Governance does three things: it lets security and platform teams set which MCP servers are approved, from an existing inventory, a curated list or servers found by discovery; it monitors MCP activity on developer machines and flags any agent reaching for a server off the list; and it can log the use for audit or block it at the moment the agent tries to connect. Snyk says enforcement runs through lightweight hooks beside the agent rather than a proxy or gateway, with policy kept centrally. It names Claude Code, Cursor, Codex and GitHub Copilot as supported today. Snyk also cites its own scan data from nearly 10,000 developer environments: 4,524 unique MCP servers in use and 1 in 12 developers with an MCP server holding a high or critical finding. Snyk's post describes the release as a first use case and says policy beyond a static allow list, such as risk-based enforcement, is still coming. These are vendor statements; Snyk gives no price in the post. ## Sources - [Evo ADS MCP Governance Goes GA, Snyk](https://snyk.io/blog/evo-ads-govern-agent-behavior-ga/) (snyk.io), read 2026-11-16 ## Related pages - [Snyk (Agent Scan, Evo ADS) profile](/vendors/snyk/) - [All news](/news/) - [MCP TypeScript SDK OAuth client advisory (CVE-2026-104850)](/news/mcp-typescript-sdk-oauth-client-credentials-advisory/) - [mcp-atlassian: 25 advisories, one critical, fixed in 0.22.0](/news/mcp-atlassian-25-advisories-authentication-bypass-fixed-0-22-0/) - [The ranking](/ranking/) --- # mcp-atlassian: 25 advisories, one critical, fixed in 0.22.0 URL: https://mcpservercybersecurity.com/news/mcp-atlassian-25-advisories-authentication-bypass-fixed-0-22-0/ Updated: 2026-09-22 GitHub published 25 advisories for the mcp-atlassian server on 22 Sep 2026: one critical (CVE-2026-77244, an HTTP transport authentication bypass), 16 high and 8 medium. Every record lists 0.22.0 as the first patched version. Teams running it over HTTP should check the installed version. 2026-09-22 The critical record, CVE-2026-77244 (GHSA-wrhw-j3f9-8vc6), says the server's HTTP transport token verifier accepted any non-empty token. The advisory describes the exposure for deployments that keep Jira or Confluence credentials in server-side environment variables, the documented quickstart pattern. It gives the vulnerable range as below 0.22.0. The other 24 records fall into groups the advisories name in their titles: reads of server-local files through attachment upload tools, server-side request forgery and incomplete earlier fixes, a toolset authorization bypass, plaintext OAuth token storage with loose file permissions, and a reflected cross-site scripting issue in the OAuth setup callback. For an engineer who inventories MCP servers, the practical step is to find every host running the pip package mcp-atlassian, confirm it is at 0.22.0 or later, and check whether the HTTP transport is reachable from anywhere other than the local machine. The count of 25 comes from the GitHub advisory database on 16 Nov 2026. ## Sources - [GHSA-wrhw-j3f9-8vc6: mcp-atlassian authentication bypass](https://github.com/advisories/GHSA-wrhw-j3f9-8vc6) (github.com), read 2026-11-16 - [NVD record for CVE-2026-77244](https://nvd.nist.gov/vuln/detail/CVE-2026-77244) (nvd.nist.gov), read 2026-11-16 ## Related pages - [All news](/news/) - [Snyk makes Evo ADS MCP Governance generally available](/news/snyk-evo-ads-mcp-governance-generally-available/) - [MCP Python SDK 2.2.0 tightens redirects and sessions](/news/python-sdk-2-2-0-redirect-session-issuer-defaults/) - [The ranking](/ranking/) --- # MCP Python SDK 2.2.0 tightens redirects and sessions URL: https://mcpservercybersecurity.com/news/python-sdk-2-2-0-redirect-session-issuer-defaults/ Updated: 2026-09-07 The MCP Python SDK published version 2.2.0 on 7 Sep 2026 with changed defaults: HTTP client redirects stay within the endpoint's origin, idle Streamable HTTP sessions expire after 30 minutes, a server holds at most 10,000 sessions, and the OAuth client checks the authorization server's issuer. 2026-09-07 The maintainers' release notes list the behavior changes first. The HTTP client follows a redirect only when it keeps the same scheme, host and port, or upgrades http to https on the same host. Any other redirect fails with an MCPError and the session stays usable. The OAuth providers apply the same rule to their own requests. For stateful servers on the legacy 2025-11-25 protocol, a session with nothing in flight for 30 minutes is closed, and a server holds at most 10,000 sessions at once. Both limits can be switched off with settings the notes name. The OAuth client now rejects authorization server metadata whose issuer is not the server's own origin on the legacy path too. The notes also add two deprecation warnings: client-credentials providers without an issuer, and server auth settings that set a resource server URL without choosing whether to validate the token's resource. A later advisory (GHSA-w4fh-qvv9-3v23, 27 Oct 2026) describes the audience gap behind the second warning, so teams running Python MCP servers should read both. ## Sources - [MCP Python SDK v2.2.0 release notes](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) (github.com), read 2026-11-16 - [GHSA-w4fh-qvv9-3v23 advisory](https://github.com/advisories/GHSA-w4fh-qvv9-3v23) (github.com), read 2026-11-16 ## Related pages - [All news](/news/) - [mcp-atlassian: 25 advisories, one critical, fixed in 0.22.0](/news/mcp-atlassian-25-advisories-authentication-bypass-fixed-0-22-0/) - [Knostic: MCP tool asks agents for a GitHub token](/news/knostic-mcp-tool-description-asks-agent-for-github-token/) - [The ranking](/ranking/) --- # MCP server security: frequently asked questions URL: https://mcpservercybersecurity.com/faq/ Updated: 2026-11-16 Short, sourced answers cover whether MCP servers are a security risk, how to secure and vet one, whether a scanner exists, how MCP differs from API security and which tools are open source. Each answer is 40 to 90 words, cites a primary source where it states a fact, and carries the read date 16 Nov 2026. ## Questions and answers ### Are MCP servers a security risk? Yes, they can be. A local MCP server is a program the client launches with the user's privileges, and the MCP specification lists arbitrary code execution, data exfiltration and data loss among the risks. Remote servers add token and network risks. Risk falls sharply when you review servers before use, sandbox them and require approval for sensitive calls. Source: [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices), read 2026-11-16. ### Is MCP server secure? Not by default. The MCP specification makes authorization optional, so a remote server with non-public tools is open until its operator adds authentication. Servers are as secure as their configuration: validated tokens, narrow scopes, localhost binding, sandboxing and logging decide the outcome, and OWASP's cheat sheet lists those controls. Source: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html), read 2026-11-16. ### How secure are MCP servers? It varies by server and by deployment. The protocol defines OAuth 2.1 based authorization and transport rules, but it leaves code review, sandboxing and monitoring to operators. A reviewed, sandboxed server with audience-checked tokens is in good shape; an unreviewed package started from a shared config file is not. Source: [MCP specification: Authorization (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization), read 2026-11-16. ### How do you secure MCP servers? Work through six steps: discover every server on every machine, vet each before first use, authenticate remote endpoints with tokens issued for your server, constrain files, network and scopes, observe with logs, and prepare to respond. The checklist on this site orders those steps, and the best practices guide explains each one. Source: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html), read 2026-11-16. ### How to secure my MCP server? Require authentication on any remote endpoint, validate the token audience, bind local servers to localhost, validate the Origin header and treat every tool argument as untrusted. The MCP transport specification requires a 403 for an invalid Origin and advises localhost binding. Add logging of each tool call with parameters. Source: [MCP specification: Transports (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/transports), read 2026-11-16. ### How to protect MCP server? Limit what it can reach and who can call it. Sandbox local servers with minimal file system and network access, use scoped per-server credentials, keep a human approval step for destructive tool calls and monitor for changed tool definitions. OWASP's MCP cheat sheet gives these controls in order of priority. Source: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html), read 2026-11-16. ### How to build a secure MCP server? Declare strict input schemas for every tool, validate paths and URLs in the handler, verify bearer tokens for your server's audience, and never forward a client's token downstream. The TypeScript and Python SDK documentation show host validation and audience settings, and our language guides walk through each. Source: [MCP TypeScript SDK (v1) documentation: Server](https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/v1.x/docs/server.md), read 2026-11-16. ### How to build secure and scalable remote MCP servers? Put authentication in front of the server, validate the audience of every token, and apply rate limits and quotas per session or tenant. Cloud services help: AWS AgentCore Runtime offers a JWT authorizer, Azure API Management validates Entra ID tokens and Cloud Run requires an IAM invoker role by default. Source: [AWS documentation: Deploy MCP servers in AgentCore Runtime](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-mcp.html), read 2026-11-16. ### What is MCP server security? MCP server security is the set of controls that decide which Model Context Protocol servers run, what they can reach and what their tools may do. It covers local servers started from config files and remote servers reached over HTTP, including authentication, sandboxing, vetting, tool-call approval and logging. Source: [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices), read 2026-11-16. ### How do you check an MCP server is safe? Read the source, the startup command and every tool description and schema before it runs, and confirm the publisher and package name. OWASP advises verifying package integrity and scanning dependencies, and the MCP specification asks clients to show the exact command before a local server starts. Re-check when definitions change. Source: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html), read 2026-11-16. ### Is there an MCP server vulnerability scanner? Yes. Cisco's open-source MCP Scanner combines YARA rules and LLM-based analysis and scans a server's tools, prompts, resources and instructions, and it can read servers from well-known client config locations. Snyk's Agent Scan finds servers across many clients. Scanners read definitions and code, so they can miss behavior hidden behind unchanged metadata. Source: [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. ### Which open source MCP security scanners exist? Two documented ones carry Apache-2.0 licenses: Cisco's MCP Scanner and Snyk's agent-scan, according to their GitHub records. OWASP's MCP cheat sheet also names mcp-scan for detecting poisoned descriptions and cross-server shadowing. Check each project's current license and maintenance before relying on it. Source: [Cisco AI Defense MCP Scanner repository](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. ### What is the difference between MCP server security and MCP gateway security? MCP server security covers the servers and the machines they run on: finding them, vetting them, authenticating them and limiting their tool calls. A gateway is a proxy that sees only traffic routed through it. This site scores tools that work on machines and does not rank gateways. ### MCP security vs API security: how do they differ? In an API, developers decide every call. In MCP a model decides which tool to call and with what arguments, and it reads the descriptions of every connected server at once. Authentication, TLS, rate limits, input validation and logging carry over from API security, per OWASP. Source: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html), read 2026-11-16. ### What is on an MCP security checklist? A useful checklist has six steps: discover servers and config files, vet source and tool definitions, authenticate with audience-checked tokens, constrain access and approvals, observe with logs and alerts, and respond with removal and revocation. The checklist on this site tracks your progress in your browser's local storage only. Source: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html), read 2026-11-16. ### What are the MCP security best practices? Run only reviewed servers, show the exact startup command before a local server runs, sandbox it, require authentication on remote endpoints, accept only tokens issued for your server, request narrow scopes, require human approval for destructive calls, pin reviewed tool definitions and log every call. The flagship guide details each one. Source: [MCP specification: Security Best Practices (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/basic/security_best_practices), read 2026-11-16. ### What does an MCP server security audit cover? An audit checks the inventory of servers, the review record for each, authentication and token audience, input validation, transport settings such as Origin checks and localhost binding, and the logs. OWASP recommends regular audits and simulated attacks against MCP setups, with logs sent to a SIEM. Source: [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html), read 2026-11-16. ## Related pages - [Guides](/guides/) - [Glossary](/glossary/) - [The ranking](/ranking/) - [Method](/method/) --- # About MCP Server Security URL: https://mcpservercybersecurity.com/about/ Updated: 2026-11-16 MCP Server Security is a reference for platform and security engineers whose developers install MCP servers. It publishes a sourced risk register, a 42-step checklist in six stages, twelve guides, a glossary, dated news and a scored comparison of 14 tools that discover and control MCP servers on machines. Sources are dated 16 Nov 2026. ## Who is this site for? The reader is a platform or security engineer whose developers install MCP servers in IDEs, desktop AI apps and command-line agents, and who must find those servers, vet them and control their tool calls. Gateways are outside the scope. The site covers what runs on machines and in developer tools, plus the hosting settings for servers the reader builds. ## What does the site publish? Two datasets lead: the MCP Risk Register and the MCP Security Checklist, each downloadable under CC BY 4.0. Twelve guides, a 28-term glossary, twelve dated news items, a FAQ and the scored comparison complete the set. Every fact carries its source URL and the date it was read. A vendor's statement about its own product is written as the vendor's. ## How is the comparison made? One research plan was applied to every vendor on 16 Nov 2026, scoring was blind, and the weights were committed before any score. The method page gives the criteria, anchors and rules in full. ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Editorial policy URL: https://mcpservercybersecurity.com/editorial-policy/ Updated: 2026-11-16 Every fact on MCP Server Security carries a source URL and a read date, and a vendor's claim is written as the vendor's. Scores follow weights committed before any evidence. Guides contain risks, detection, configuration and fixes only, with no exploit code, payloads or bypass steps. Changes follow a sourced fact or a reported error. ## How are facts sourced? Facts come from primary pages: the MCP specification, OWASP, GitHub and CVE records, cloud and client documentation, and the vendor's own pages and release notes. Each stored fact keeps the quoted passage, the URL, the read date and its kind: vendor, third-party, observed or none found. Where evidence is missing the page says none found on the read date. It does not infer a capability from silence. ## How are vendor claims handled? A vendor's claim is attributed, for example Zenity says or per Microsoft's documentation. A claim alone earns at most 7 of 10 on any criterion, and a higher score needs a second source such as third-party reporting or something observed in public documentation or code. Each vendor's strength is stated fairly, and each page lists the gaps the documentation shows. ## What will the guides not contain? Guides describe risks, how to detect them, configuration and fixes. They carry no exploit code, no prompt-injection payloads and no steps to bypass a control. ## When do pages change? A score changes only when a sourced fact changes, and the change is logged with the fact. Weights and criteria stay fixed after the first score. Corrections are described on the corrections page. ## Related pages - [Method](/method/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Corrections and tips URL: https://mcpservercybersecurity.com/corrections/ Updated: 2026-11-16 To report an error, email editor@mcpservercybersecurity.com with the page address, the sentence in question and a primary source that contradicts it. Corrections are checked against that source and logged. The same address takes tips on MCP vulnerabilities, malicious packages and product changes that a page should cover. ## How do I report an error? Send the page address, the sentence you believe is wrong and a link to the primary source that shows the correct fact. Vendors may also send documentation that was not public on 16 Nov 2026. Write to editor@mcpservercybersecurity.com (mailto:editor@mcpservercybersecurity.com). ## How are corrections handled? Each report is checked against the source you supply and against the page the original fact came from. If the correction stands, the fact is updated, the score is recomputed only where the fact feeds a score, and the change is logged on the method page with the date. If a report cannot be confirmed from a public source, the page stays as published and the reason is recorded. ## How do I send a tip? Send a tip to the same address with the subject line Tip. Useful tips name a vulnerability with its CVE or advisory number, a malicious or typosquatted package with its registry link, a specification change, or a release note from a tool the site scores. Tips are read against primary sources before anything is published, and a tip alone is never printed as fact. ## Write to the editor Email [editor@mcpservercybersecurity.com](mailto:editor@mcpservercybersecurity.com) with the page address, what is wrong and the source that shows it. ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Contact MCP Server Security URL: https://mcpservercybersecurity.com/contact/ Updated: 2026-11-16 Email editor@mcpservercybersecurity.com with questions about the register, the checklist, a guide or a score, or with a correction or a tip. Put the topic in the subject line, include the page address, and expect a reply to the address you wrote from. ## How do I contact the editor? Use editor@mcpservercybersecurity.com for every request. Write to editor@mcpservercybersecurity.com (mailto:editor@mcpservercybersecurity.com). ## Which subject lines help? Messages that include the page address and a source link are answered fastest. - Question: about a score, a method rule or a guide. - Correction: an error on a page, with the page address and a primary source. - Tip: an MCP vulnerability, package or specification change, with a link. - Data: use of the CC BY 4.0 datasets, with the edition date you cite. ## Write to the editor Email [editor@mcpservercybersecurity.com](mailto:editor@mcpservercybersecurity.com). ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Privacy policy URL: https://mcpservercybersecurity.com/privacy/ Updated: 2026-11-16 MCP Server Security is a static site. CloudFront access logs are kept for 90 days. The site sets no cookies and runs no analytics. The checklist may keep your progress in your browser's local storage only, and that data never leaves your device. Email you send to the editor is used to reply. ## What data does the site collect? Pages are served as static files through CloudFront. The access logs record the requested path, time, response status, user agent and IP address, and they are kept for 90 days, then deleted. The logs are used to keep the site running and to find errors. They are not combined with any other data. ## Does the site use cookies or analytics? No. The site sets no cookies and loads no analytics, advertising or tracking scripts. ## What does the checklist store? The checklist can keep your ticked items in your browser's local storage so progress survives a reload. That data stays on your device, is not sent to any server and can be cleared in your browser's site-data settings. ## What happens to email? If you write to editor@mcpservercybersecurity.com, your address and message are used to reply and, for corrections, to record what changed. Ask in the same thread if you want the message deleted. ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Terms of use URL: https://mcpservercybersecurity.com/terms/ Updated: 2026-11-16 MCP Server Security publishes its two datasets under the CC BY 4.0 license, and the cite line gives the asset name, the edition date of 16 Nov 2026, the page URL and the license. Each fact carries its read date and a link to its source. Errors are reported through the corrections page. ## What may I reuse? The two datasets are licensed under CC BY 4.0. Cite them as: MCP Server Security, the asset name, edition 16 Nov 2026, the page URL, CC BY 4.0. Other text may be quoted briefly with a link to the page. ## How is the information dated and sourced? Every fact is printed with the read date beside its source link. Scores come from public documentation, code and third-party reporting read on that date. A fact that changes is corrected and logged on the method page. ## How do I report a problem? Use the corrections page. Reports with a primary source are checked and, where confirmed, logged and fixed. ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/)