mcp-builder

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

• Third-party content exposure detected (high risk: 0.90). The skill's workflow (SKILL.md) explicitly instructs the agent to fetch and read public web resources—e.g., https://modelcontextprotocol.io/sitemap.xml and raw GitHub READMEs like https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/README.md—and to "use web search and WebFetch as needed", meaning untrusted third-party content is ingested and used to guide tool design and agent actions, allowing indirect prompt injection.

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 runtime MCP server URL (e.g., https://example.com/mcp) and uses it to fetch tool definitions via SSE/HTTP which are passed into the model as "tools" (controlling prompts/available actions) and can trigger remote tool execution, so the remote URL's content directly controls agent behavior at runtime.