cisco-ai-defense/skill-scanner is a security scanner that inspects AI agent skill packages before they are installed. It checks the code and documents bundled in a skill with static analysis and LLM analysis to surface risks, and it prints CI results. It is implemented in Python, and the latest release in this repository is 2.3.0. The collected material does not contain the change log or the release date for 2.3.0.
The README describes the scanner as a best-effort security inspection tool that finds prompt injection, data exfiltration, and malicious code patterns in agent skills. Its analysis layers are pattern analysis based on YAML and YARA-X, AST (a tree that represents the structure of code) and data flow analysis, an optional LLM-as-a-judge (the model judges the risk itself), and a scoped CEL (Common Expression Language) decision layer. It has 2616 GitHub stars, and the repository was created on January 29, 2026.
The README also states that the results do not guarantee security even when nothing is found, and that human review is required. There are no benchmarks that gauge detection rate or false positive rate. The license is listed as NOASSERTION on GitHub, so the actual distribution terms have to be confirmed separately. The only recorded score is Momentum at 2.0 out of 10, and no other metric is computed. It can work as a first-pass screening tool before installation, but it is still too early to treat its results as a basis for trust.