eval-and-improve

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

• Third-party content exposure detected (medium risk: 0.65). The required runtime workflow (python -m evals via evals/__main__.py and evals/cases.py) runs LLM judge evals (AgentAsJudgeEval) and may invoke a web-search tool whose results are fetched from public web content at runtime and then ingested as readable text into the agent/judge context, creating an outsider free-text injection surface.