mcp-gateway

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.75). This skill connects to external MCP servers via HTTP/SSE/WebSocket/stdio and reads their JSON-RPC responses (including free-text fields like tool descriptions/error messages) at runtime, which can be outsider-authored and then supplied into the agent’s LLM context through tool-call results.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The configuration runs "npx" with the package "@modelcontextprotocol/server-filesystem" at runtime (SKILL.md:813-814), which will fetch and execute remote code from the npm registry, so this is a runtime external dependency that executes code.