skill-safety-scanner

name	skill-safety-scanner
description	Run local safety scans on Agent Skills before publishing. Detects secrets, dangerous code patterns, and analyzes required permissions.
version	1.0.0
license	MIT
author	@anthropic-agent-skills
tags	["safety","security","scanning","validation"]

Instructions

Use this skill to run safety scans on any Agent Skill directory before publishing. The scanner detects:

• Secrets - Hardcoded API keys, tokens, passwords, private keys
• Dangerous Code - eval(), exec(), command injection, XSS patterns
• Permissions - Required capabilities (filesystem, network, subprocess, etc.)

When to Use

• Before submitting a skill for publication
• To preview what the catalog safety scan will find
• To identify and fix security issues early
• As part of CI/CD pipelines

How to Use

Scan a skill directory:

Scan the skill at /path/to/my-skill for safety issues

Get detailed output:

python3 safety_scan.py /path/to/my-skill --verbose

Get JSON output for CI integration:

python3 safety_scan.py /path/to/my-skill --json

Output

Safety Scan Report
  Skill: my-skill
  Grade: B (85/100)

  Scores:
    Secrets: 100/100
    Dangerous Code: 70/100

  Permissions Detected:
    - filesystem
    - network

  Findings (1):
    [medium] Potential command injection
      File: scripts/main.py:42
      Code: subprocess.run(cmd, shell=True)

  Recommendation: Review the finding above before publishing.

Examples

Basic scan:

User: Scan my skill at ./document-tools for safety issues
Agent: Running safety scan on ./document-tools...

       Grade: A (95/100)
       No critical issues found.

Finding secrets:

User: Check ./my-api-client for security issues
Agent: Running safety scan...

       Grade: F (0/100)

       Findings:
         [critical] Hardcoded API key detected
           File: config.py:5
           Code: API_KEY = "sk-ant-..."

       You must remove this secret before publishing.

Limitations

• Does not scan dependencies (use npm audit / pip-audit separately)
• Pattern-based detection may have false positives
• Cannot detect all security issues (not a replacement for security review)
• Scans local files only (not GitHub URLs)

Dependencies

• Python 3.9+
• No external dependencies (uses Python stdlib)