SkillCheck for Claude Skills

Security scanner for Claude SKILL.md files. Detects vulnerabilities before they reach production.

Features

• Hardcoded Secrets Detection - Finds API keys, tokens, passwords, and credentials
• Command Injection Detection - Identifies unsafe shell command execution
• Path Traversal Detection - Catches directory traversal vulnerabilities
• Smart False Positive Reduction - Context-aware analysis and placeholder detection
• Beautiful CLI Output - Color-coded, actionable security reports

Installation

npm install @agentigy/skillcheck

Or for development:

npm install
npm run build

Usage

Scan a single file

npx skillcheck .claude/skills/my-skill.md

Scan a directory

npx skillcheck .

Configure failure threshold

# Fail CI on HIGH or CRITICAL issues (default: CRITICAL)
npx skillcheck --fail-on HIGH ./skills

Generate SARIF output for GitHub Security

# Output SARIF format for GitHub Code Scanning
npx skillcheck --format sarif . > results.sarif

Security Rules

CRITICAL Severity

Hardcoded Secrets (SECRET_EXPOSURE_001)
- Detects: AWS keys, API keys, JWT tokens, private keys, GitHub tokens
- Excludes: Placeholders like your_api_key_here, ${API_KEY}, XXXX
- CWE-798

Command Injection (CMD_INJECTION_001)
- Detects: Unsafe bash, eval(), exec(), os.system() with user input
- Context-aware: Skips when validation code is present nearby
- CWE-78

Privilege Escalation (PRIV_ESCALATION_001)
- Detects: sudo, setuid/setgid, chmod 777, su, pkexec, doas, kernel module loading, Docker socket access, system file modifications
- Context-aware: Skips when authorization checks are present nearby
- CWE-250

HIGH Severity

Path Traversal (PATH_TRAVERSAL_001)
- Detects: ../ sequences, file operations with unvalidated user paths
- Context-aware: Skips when path.resolve() or validation exists
- CWE-22

Information Disclosure (INFO_DISCLOSURE_001)
- Detects: SSH keys, AWS credentials, .env files, /etc/passwd, environment variables, database dumps, browser cookies, Git credentials, connection strings, process lists
- Context-aware: Skips when redaction/filtering code is present nearby
- CWE-200

Output Formats

Console (default)

Human-readable output with color-coded severity levels. Best for local development and interactive use.

npx skillcheck ./skills

SARIF

Machine-readable JSON format compatible with GitHub Code Scanning and other security platforms. Use this for CI/CD integration.

npx skillcheck --format sarif . > results.sarif

You can upload SARIF results to GitHub Security tab using:

# .github/workflows/security.yml
- name: Upload SARIF results
  uses: github/codeql-action/upload-sarif@v2
  with:
    sarif_file: results.sarif

Exit Codes

• 0 - No issues found (or below fail threshold)
• 1 - Security issues found at or above fail threshold

Development

Run tests

npm test

Run tests in watch mode

npm run test:watch

Build

npm run build

Lint

npm run lint

Examples

Vulnerable Skill (Will Fail)

# Database Connector

Connect using:
api_key = "sk_live_51HqT2jK3xR9pQ8vN4mL5wY2"

Execute command:
os.system(f"rm {userInput}")

Output:

❌ Critical security issues found!
- Potential Generic API Key detected (Line 4)
- Potential command injection (Line 7)

Secure Skill (Will Pass)

# Database Connector

Validate input first:
if userInput not in ALLOWLIST:
    raise ValueError("Invalid input")

Connect using environment variable:
api_key = os.getenv("API_KEY")

Execute with validated input:
os.system(f"rm {shlex.quote(userInput)}")

Output:

✓ No security issues found!

Architecture

• Parser - Extracts structure from skill.md files, identifies code blocks
• Scanner - Orchestrates rule execution, filters false positives
• Rules - Modular security checks with context-aware analysis
• Reporters - Multiple output formats (console, SARIF)
• CLI - Command-line interface for local and CI usage

Future Enhancements

• [x] SARIF output format for GitHub Security tab
• [x] Privilege escalation detection
• [x] Information disclosure detection
• [ ] JSON output format
• [ ] GitHub Actions workflow
• [ ] Configuration file support
• [ ] Custom rule definitions
• [ ] VS Code extension

Built by Agentigy — securing AI agents and workflows.

License

MIT