skill-scanner

Why we included this project

Agent skills are just files, and when you pull one in from a public repo you are effectively running someone else's instructions inside your agent, which is why a skill can quietly hide prompt injection or malicious code that exfiltrates your data. Skill Scanner checks those files before they execute, layering YAML and YARA signature matching over LLM-based semantic review and behavioral dataflow analysis so no single detection method carries the whole load. The project is upfront that this is best-effort scanning, not a guarantee: a clean result does not mean a skill is safe, and human review still has a place. That honesty is the right tone for a security tool. It also drops into existing workflows with a CLI, a REST API, SARIF output for GitHub Code Scanning, and a reusable GitHub Actions workflow, and the plugin architecture lets you add your own analyzers or rule packs. If you build or consume agent skills and want a defense-in-depth check rather than a false sense of certainty, this is a reasonable place to start.