Skill Vetter - Pre-Install Security Review

**Quick judgment** Strongly recommended as a baseline, security-first vetting protocol before installing any AI agent skill, especially from unverified or low-trust sources. It is conservative by design, and will tend to favor safety over convenience. **What it does well** - Provides a concrete, stepwise vetting process (source check → code review → permission scope → risk classification). - Defines clear “REJECT IMMEDIATELY” red flags (credential/file theft patterns, exfiltration endpoints, base64+eval chains, obfuscated/minified code, unpinned or hidden dependencies, sudo/root requests, etc.). - Enforces the principle of least privilege by explicitly reviewing file, network, and command scopes. - Supplies practical templates: a checkbox-based vetting checklist and a structured “Skill Vetting Report” format for consistent documentation. - Differentiates risk levels (LOW/MEDIUM/HIGH/EXTREME) with example scenarios and recommended actions. **Key risks / limitations** - This is a process/template skill, not an automated static analyzer; effectiveness depends entirely on the thoroughness and competence of the reviewer. - Highly conservative rules may yield false positives, potentially rejecting some legitimate but complex skills (e.g., those legitimately using external APIs, base64, or certain shell patterns). - No built-in enforcement or sandboxing; it guides human/agent judgment rather than actively constraining what a skill can do after installation. - Requires full code access; closed-source or partially hidden skills cannot be vetted properly under this protocol. **Recommended scenarios** - Mandatory use before installing any new or untrusted ClawHub or GitHub skill, particularly those touching credentials, trading, system configuration, or wide filesystem access. - As part of an organization’s standard security policy for agent ecosystems, especially in regulated or sensitive environments. - For documenting and auditing installation decisions, where security teams want a paper trail of how each skill was evaluated and classified.