AI Skill Security Vetter

Problem

AI agent skills often come from ClawdHub, GitHub, or other agents, where source, dependencies, and permission scope may be opaque. Installing unknown code can expose credentials, trigger system commands, trading actions, or browser automation. Skill Vetter turns install-or-not into a reviewable security question rather than a judgment based on author trust.

Workflow

The skill structures vetting into concrete steps:
- Source check: determine whether the skill comes from an official source, a high-star repository, a known author, or an unknown source.
- Code review: read all files in the skill and check for RED FLAGS, not just the entrypoint script.
- Permission scope: identify file operations, browser control, APIs, credentials, and system-level capabilities.
- Risk classification: map notes, weather, or formatting tools to low risk; credentials, trading, or system access to high risk; security configs or root access to extreme risk.
- Vetting report: produce a structured conclusion covering install OK, full review required, human approval required, or do not install.

Boundaries

It emphasizes “when in doubt, don’t install” and requires human approval for high-risk decisions. Trust hierarchy still reviews all sources, even official or high-star ones. It does not replace sandboxing, signing, runtime isolation, or security audits; it is a pre-install behavioral review protocol.