Skill Vetter Pro

Problem

When an agent installs a third-party skill, the risk is often hidden behavior: unsafe code, file overreach, outbound transmission, vulnerable dependencies, or post-install tampering. skill-vetter-pro turns ad hoc review into a repeatable, traceable security workflow.

How It Works

It is organized around a four-dimensional review. skill_vet.py uses AST scanning to detect patterns such as eval, exec, base64 decoding, and SSH key reads. skill_behavior_monitor.py uses strace to observe system calls, file changes, process creation, and network activity. skill_integrity_checker.py verifies SHA256 state to reduce post-review tampering risk, while dependency_cve_checker.py audits dependencies for known vulnerabilities.

The flow confirms the target and policy, checks Docker, images, and scripts, runs the review in a sub-session via sessions_spawn, and renders a verdict as LOW / MEDIUM / HIGH / EXTREME before logging it to records/vetting-log.md.

Boundaries

Use it before installing third-party skills, when re-reviewing official skills, or after alerts. Static scanning only detects known risky patterns. Runtime monitoring depends on the sandbox. CVE checks cover common manifests but do not replace patching.