MCP Server Security

MCP execution and filesystem risks

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.

Facts read on .

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; MCP Security Best Practices, 2026-07-28 revision; Connect Claude Code to tools via MCP, Claude Code docs.

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.

Sources: Local Server Security, modelcontextprotocol.io; MCP Security Best Practices, 2026-07-28 revision; Connect Claude Code to tools via MCP, Claude Code docs; MCP Inspector 2.12.0 release notes.

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.

Sources: Local Server Security, modelcontextprotocol.io; MCP Security Best Practices, 2026-07-28 revision; Connect Claude Code to tools via MCP, Claude Code docs; MCP Inspector 2.12.0 release notes.

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 (read ); MCP Security Best Practices, 2026-07-28 revision (read ).

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 (read ); Connect Claude Code to tools via MCP, Claude Code docs (read ).

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 (read ); MCP Inspector 2.12.0 release notes (read ); OWASP MCP Security Cheat Sheet (read ).

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 (read ); OWASP MCP Security Cheat Sheet (read ).

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 (read ); AI settings for enterprise, Visual Studio Code docs (read ).

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 (read ); MCP Tools, 2026-07-28 revision (read ).

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 (read ); MCP Inspector 2.12.0 release notes (read ).