mcp-builder

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

• Third-party content exposure detected (high risk: 0.90). The SKILL.md workflow (Phase 1.2 and the "Reference Files" sections) explicitly instructs the agent to fetch and read public web pages (e.g., https://modelcontextprotocol.io/sitemap.xml and https://raw.githubusercontent.com/.../README.md) and to "use web search and WebFetch as needed", which are untrusted third-party sources the agent is expected to read and that can materially influence tool design and subsequent actions.

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

• Potentially malicious external URL detected (high risk: 0.90). The evaluation harness accepts a user-specified MCP server URL (e.g., https://example.com/mcp) and at runtime connects to that endpoint to call remote tools whose execution and outputs are injected back into the agent, meaning the external URL can execute code and directly control agent prompts/results.