mcp-builder

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

• Third-party content exposure detected (high risk: 0.75). The required runtime workflow in scripts/evaluation.py feeds tool outputs back into the agent’s LLM context via messages.append(... "type": "tool_result" ... content: tool_response ...), and those tool results can include outsider-authored free text from an MCP server connected over stdio/SSE/HTTP (i.e., not authored by the operating user).

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly instructs the agent to WebFetch external documentation at runtime (e.g., https://modelcontextprotocol.io/llms-full.txt and the two raw.githubusercontent.com README URLs), which will be fetched during execution and used as authoritative instructions/specifications for the agent.