eval-engineering

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

• Third-party content exposure detected (medium risk: 0.30). The skill’s multi-turn simulation runtime can ingest outsider-authored free text when the user selects the “LLM user” simulation track, because run_llm_user_conversation feeds the simulated user’s generated message back into the Harness session as user input, and that simulated user text is produced from the (user-provided) contract/transcript/observation payload.