skill-eval-improve

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

• Third-party content exposure detected (medium risk: 0.30). In SKILL.md’s Layer 2/3/4 workflow, the skill-eval-improve process consumes outsider-authored free text from the evals/cases/*.yaml input field (e.g., attacker-provided prompts during pnpm run eval) and feeds it to the agent to decide improvements.