skill-vetter

Vetting AI skills for security risks involves a four-step process: checking the source reputation, reviewing code for red flags, evaluating permission scopes, and classifying the overall risk level to output a safety verdict.

Red flags in AI agent skills from GitHub repositories include suspicious code patterns, unnecessary read/write access, broad execution privileges, and unsafe network access that exceed the minimal necessary permission scope.

Permission scope evaluation for external AI skills assesses requested read/write access, network access, and execution privileges to ensure the skill only has the minimal necessary permissions required to function safely.

Yes, you can apply this security-first vetting protocol to evaluate AI skills from any external source, including GitHub repositories and marketplaces, to identify red flags and prevent unsafe deployments prior to installation.

The best way to classify the risk level of an AI agent skill is to complete source, code, and permission checks, then assign a risk classification that outputs a clear verdict with recommended mitigations.

You should not install an AI skill from an unverified author when the vetting process identifies unsafe code patterns, excessive permission scopes, or a high-risk classification that lacks clear mitigations.