skill-audit

🔒 Security auditing CLI for AI agent skills.

AI agents increasingly use "skills" - modular packages containing instructions and executable code that extend an agent's capabilities. These skills can execute code on your machine, making them a significant security risk if not properly vetted.

skill-audit helps you audit skills before trusting them.

Features

• 🔍 Prompt Injection Detection - Scans skill descriptions for jailbreak patterns and manipulation attempts
• 🔑 Secret Scanning - Finds hardcoded API keys, tokens, and credentials (via trufflehog/gitleaks)
• 🐚 Shell Script Analysis - Checks bash scripts for dangerous patterns (via shellcheck)
• 🐍 Code Security - Analyzes Python/JS code for security issues (via semgrep)
• 📄 SARIF Output - CI/CD ready output format for GitHub Actions integration
• 🔌 Extensible - Plugin architecture for custom scanners

Installation

macOS

# 1. Clone the repo
git clone https://github.com/markpors/skill-audit
cd skill-audit

# 2. Create virtual environment and install
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

# 3. Install security tools (recommended)
brew install shellcheck semgrep trufflehog

# Note: gitleaks is an alternative to trufflehog (only need one)
# brew install gitleaks

# 4. Verify installation
skill-audit check-tools

Add to your shell profile (optional, for global access):

# Add to ~/.zshrc or ~/.bashrc
alias skill-audit="source ~/dev/skill-audit/.venv/bin/activate && skill-audit"

Linux

# 1. Clone and install
git clone https://github.com/markpors/skill-audit
cd skill-audit
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

# 2. Install security tools
pip install semgrep
sudo apt install shellcheck  # Debian/Ubuntu
# For trufflehog: https://github.com/trufflesecurity/trufflehog#installation

From PyPI

pip install skill-audit

Usage

Basic Audit

# Audit a skill directory
skill-audit audit ./my-skill/

# Audit a specific file
skill-audit audit ./my-skill/SKILL.md

Output Formats

# Pretty terminal output (default)
skill-audit audit ./my-skill/

# JSON output
skill-audit audit ./my-skill/ --format json

# SARIF output (for CI/CD)
skill-audit audit ./my-skill/ --format sarif -o results.sarif

Check Available Tools

skill-audit check-tools

Strict Mode

# Fail on warnings (not just errors)
skill-audit audit ./my-skill/ --strict

What It Checks

Prompt/Instruction Analysis

• Jailbreak patterns ("ignore previous instructions", "DAN mode", etc.)
• Role manipulation attempts ("you are root", "pretend to be admin")
• Data exfiltration instructions
• Safety bypass attempts

Code Analysis

• Hardcoded secrets and credentials
• Dangerous shell patterns (rm -rf, eval, etc.)
• Arbitrary code execution risks
• Subprocess injection vulnerabilities

Exit Codes

Code	Meaning
0	Passed - no errors found
1	Failed - errors found (or warnings in strict mode)
2	Error - tool execution failed

CI/CD Integration

GitHub Actions

- name: Audit Skills
  run: |
    pip install skill-audit
    skill-audit audit ./skills/ --format sarif -o results.sarif
    
- name: Upload SARIF
  uses: github/codeql-action/upload-sarif@v2
  with:
    sarif_file: results.sarif

Development

# Clone and install in dev mode
git clone https://github.com/markpors/skill-audit
cd skill-audit
pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check src/

Limitations

skill-audit is a static analysis tool. It cannot catch everything.

What it CAN detect:

• Known jailbreak patterns and prompt injection attempts
• Hardcoded secrets and credentials
• Dangerous code patterns (eval, exec, shell injection)
• Common shell script vulnerabilities

What it CANNOT detect:

• Obfuscated malware — encoded, encrypted, or cleverly hidden payloads
• Novel attack patterns — zero-day techniques not in our ruleset
• Contextual intent — a "File Deleter" skill legitimately needs to delete files
• Indirect prompt injection — malicious content injected via external data sources
• Runtime behavior — dynamic code generation, network calls at runtime
• Supply chain attacks — compromised dependencies of the skill itself

False positives

Some legitimate skills may trigger warnings. For example:

• A password manager skill will "access credentials" — that's its job
• A cleanup utility will use rm commands
• An API client will contain endpoint URLs

Use --strict mode thoughtfully, and review warnings in context.

Disclaimer

⚠️ THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.

skill-audit is a security scanning tool, not a guarantee of safety.

• A passing audit does not mean a skill is safe. It only means no known issues were detected.
• A failing audit does not mean a skill is malicious. It may be a false positive.
• Always review skills manually before granting them access to sensitive systems.
• Do not rely solely on this tool for security decisions in production environments.
• The authors are not responsible for any damage caused by skills that pass or fail audits.

Use at your own risk. When in doubt, don't install the skill.

License

MIT