ai-evals

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

• Third-party content exposure detected (medium risk: 0.65). The skill’s runtime LLM context is built from evaluation inputs/outputs (e.g., query, response, and optional context) that come from the eval dataset loaded from storage (e.g., s3://evals/datasets/.../*.jsonl.gz), which can include outsider-authored free text such as production-sampled user queries or adversarial dataset content; that text is then interpolated into the judge prompt (Query: {query} Response: {response} Context: {context}) and fed to the judge LLM.