MCP supply chain risks
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.
Facts read on .
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; Local Server Security, modelcontextprotocol.io; Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic.
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; Local Server Security, modelcontextprotocol.io; Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic; MCP Security Best Practices, 2026-07-28 revision.
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; Local Server Security, modelcontextprotocol.io; Repackaged Amazon SP-API MCP Sends App Secrets to an Undisclosed IP, Knostic; MCP Security Best Practices, 2026-07-28 revision.
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 (read ); Local Server Security, modelcontextprotocol.io (read ).
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 (read ); Local Server Security, modelcontextprotocol.io (read ).
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 (read ); Local Server Security, modelcontextprotocol.io (read ).
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 (read ); Evo ADS MCP Governance Goes GA, Snyk (read ).
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 (read ); MCP Tools, 2026-07-28 revision (read ).
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 (read ); MCP Registry v1.8.8 release notes (read ).
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 (read ); What's new in Microsoft Defender for Endpoint (read ).