skscan

Security scanner for AI agent skills — detect secrets, prompt injections, dangerous code, data exfiltration, and hidden instructions in SKILL.md files.

Install

npx skscan .          # run without installing
npm install -g skscan # or install globally

Scan

skscan .              # scan current directory
skscan ./my-skill/    # scan a directory
skscan ./SKILL.md     # scan a single file

Options

-f, --format <fmt>   pretty | json | sarif (default: pretty)
-s, --strict         exit 1 on any finding
--ignore <rules>     comma-separated rule IDs to skip
-c, --config <path>  path to config file
--badge              output SVG badge to stdout

Config

Create .skscanrc.json:

{
  "rules": {
    "secrets/high-entropy": "off"
  },
  "ignore": ["node_modules/**", "dist/**"]
}

Or run skscan init to generate one.

CI

# .github/workflows/scan.yml
name: Security Scan
on: [push, pull_request]
jobs:
  skscan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npx skscan ci .

The ci command outputs JSON, emits GitHub annotations, and exits 1 on any finding.

Exit Codes

| Code | Meaning | |------|---------| | 0 | Pass | | 1 | Fail (findings detected) | | 2 | Error |

29 Rules, 5 Categories

• Secrets (7) — AWS keys, GitHub tokens, private keys, passwords, high-entropy strings
• Dangerous Code (8) — Remote code execution, destructive commands, dynamic execution
• Prompt Override (6) — Instruction hijacking, role reassignment, restriction removal
• Exfiltration (4) — Environment variable extraction, sensitive path access
• Hidden Instructions (4) — Zero-width chars, invisible unicode, HTML comment injection

Full rules reference and API docs at github.com/Khaledgarbaya/skillvault.

License

MIT