recipe-eval-skill

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

• Third-party content exposure detected (high risk: 0.75). Held-out requests (Round 3) are verbatim user-provided free text that Phase B feeds into eval-executor.py via the claude -p --prompt runtime path, and this is outsider text because it comes from the operating user’s provided team requests/comments rather than the skill author’s own authored material.