cisco-ai-defense/skill-scanner

What it solves

Skill Scanner is a security tool designed to detect threats in AI Agent Skills (such as OpenAI Codex and Cursor Agent Skills). It specifically targets risks like prompt injection, data exfiltration, and malicious code patterns that could be introduced into agentic workflows.

How it works

The tool uses a multi-engine detection approach to maximize coverage and reduce false positives:

• Static Analysis: Uses YAML and YARA patterns to scan files.
• Behavioral Analysis: Employs AST dataflow analysis for Python files and command taint analysis for shell pipelines.
• Semantic Analysis: Uses LLMs (via API) to analyze the meaning of SKILL.md files and scripts.
• Meta-Analysis: A dedicated meta-analyzer filters out noise and prioritizes findings to reduce false positives.
• External Integration: Integrates with VirusTotal for binary hash scanning and Cisco AI Defense for cloud-based AI analysis.

Who it’s for

Developers and security engineers who build, deploy, or integrate AI agents and want to ensure the skills they are using are not malicious or vulnerable.

Highlights

• Multi-Engine Detection: Combines static, behavioral, LLM-based, and cloud scanning.
• CI/CD Integration: Supports SARIF output for GitHub Code Scanning and provides a reusable GitHub Actions workflow.
• Pre-commit Hook: Integrates with the pre-commit framework to scan skills before they are committed to version control.
• Extensible Architecture: Features a plugin system for adding custom analyzers and custom YARA rules.
• Interactive Setup: Includes a CLI wizard and a TUI for policy configuration.