mcp-builder

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

• Third-party content exposure detected (medium risk: 0.30). The evaluation harness reads the question field from the outsider-authored evaluation.xml at runtime and passes it directly as the user message to Claude (scripts/evaluation.py → parse_evaluation_file() → agent_loop() → client.messages.create()), so free text from the evaluation file can reach the LLM before selecting any specific tool item.

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

• Potentially malicious external URL detected (high risk: 0.90). The evaluation harness can connect to an external MCP server URL (e.g., https://example.com/mcp) at runtime and loads that server's tool definitions, which are passed to the LLM as the "tools" parameter and therefore can directly control agent behavior.