trace-to-training-data

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

• Third-party content exposure detected (high risk: 0.70). SKILL.md’s required workflow ingests graded trace JSONL from eval/goldens.jsonl and runs/<run-id>/results.json (which contain messages authored by the rollout/user/assistant), and those messages are then converted into training rows—so the LLM context can include outsider-authored free text via the runtime trace-to-training conversion path.