skill-scanner

README documents CI badges, a plugin architecture, typed CEL decision layer using cel-go v0.32.0, and detailed benchmark methodology with precision/recall/F1 metrics.

The tool itself is a defensive scanner from Cisco AI Defense with no download-and-execute patterns shown; it transparently discloses false-positive/negative risks and keeps CEL rules in shadow mode pending promotion gates.

Addresses a real and growing pain point by detecting prompt injection, data exfiltration, and malicious patterns in agent skills, with SARIF output, pre-commit hooks, and GitHub Actions integration for CI/CD workflows.

Last updated 2026-09-05, not archived, 2516 stars, active CI workflows, Discord community, and documented ongoing detection evaluation and rollout work.

README provides install command, quick-start guide, architecture docs, custom policy configuration, GitHub Actions docs, and an explicit Scope and Limitations section with benchmark evidence.