mcp-builder

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

• Third-party content exposure detected (medium risk: 0.30). In scripts/evaluation.py, the evaluation harness ingests the XML-provided <question> free text at parse_evaluation_file() → agent_loop() as the user prompt (messages = [{"role": "user", "content": question}]) without any filtering, and that prompt is then used to drive tool calls against the MCP server.