improve-skills

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

• Third-party content exposure detected (high risk: 0.70). The workflow runs eval_env inspect ... which executes the agent in solvers/inspect-swe against asta-bench-private cases; those cases are sourced from a separate repo and can include outsider-authored free-form text (e.g., issue/thread-style descriptions) that the agent reads and sends into the LLM context during evaluation.