Skill Vetting

**Judgement:** Strongly recommended as a security-focused helper skill for vetting other skills, provided it is treated as an aid to human judgment rather than a final authority. **What it does:** - Downloads skills to a temporary directory, runs a Python scanner to flag suspicious patterns, and guides a structured manual code review. - Focuses heavily on detecting and resisting prompt-injection and social-engineering text aimed at AI/LLM reviewers. - Encourages evaluating both **security** (malicious patterns, prompt injection, unexpected I/O) and **utility** (does this skill provide unique value?). **Key risks / limitations:** - Scanner is regex-based and can be bypassed; it will miss more sophisticated or obfuscated attacks. - False negatives are possible; it must not be used as a sole security gate. - Requires users to actually follow the workflow (review findings in context, compare to documentation) for real protection. **Recommended scenarios:** - Before installing any new or untrusted ClawHub skill, especially those with file/network access. - When reviewing third-party code bundles where prompt-injection or social engineering against AI agents is a concern. - As part of a standard security review process in teams that use OpenClaw in production or sensitive environments.