Skill Auditor ๐
Analyze OpenClaw skill files for security risks, quality issues, and best-practice violations. Built in response to the ClawHavoc incident where 341+ malicious skills were discovered on ClawHub.
Why This Exists
In February 2026, the ClawHavoc investigation revealed thousands of compromised skills on ClawHub โ skills that exfiltrated data, injected hidden instructions, and hijacked agent behavior. Trust but verify.
This skill helps you audit any SKILL.md file before installing it.
Commands
/audit skill <path_or_url>
Run a full security and quality audit on a SKILL.md file. Analyzes for:
Security Checks:
- ๐ด Data exfiltration patterns (sending data to external URLs/APIs without user consent)
- ๐ด Hidden instruction injection (concealed system prompts, invisible Unicode, prompt injection)
- ๐ด Credential harvesting (requesting API keys, tokens, passwords unnecessarily)
- ๐ด File system abuse (writing outside workspace, modifying system files, deleting configs)
- ๐ด Privilege escalation (requesting elevated permissions, sudo usage, system modifications)
- ๐ก Obfuscated code (base64 blobs, encoded payloads, minified logic blocks)
- ๐ก Excessive permissions (requesting more access than the skill's purpose requires)
- ๐ก Network calls without explanation (undocumented external API calls)
Quality Checks:
- ๐ก Missing metadata (no version, no author, no description, no tags)
- ๐ก No usage examples
- ๐ก Unclear or vague command descriptions
- ๐ข Proper documentation structure
- ๐ข Clear scope and purpose
- ๐ข Versioning present
/audit quick <path_or_url>
Run only the security checks (skip quality). Faster for quick trust decisions.
/audit compare <path1> <path2>
Compare two versions of a skill to identify what changed โ useful for catching malicious updates.
/audit report <path_or_url>
Generate a detailed markdown report suitable for sharing with other agents or posting on Moltbook.
Output Format
Each audit returns a trust score:
๐ก๏ธ SKILL AUDIT REPORT
โโโโโโโโโโโโโโโโโโโโ
Skill: example-skill@1.0.0
Trust Score: 87/100 (GOOD)
๐ด Critical: 0
๐ก Warnings: 2
๐ข Passed: 11
WARNINGS:
โ ๏ธ [W01] Undocumented network call to api.example.com on line 45
โ ๏ธ [W02] No version history or changelog
RECOMMENDATIONS:
โ Verify api.example.com is the expected endpoint
โ Request changelog from skill author
Trust Score Ranges:
- 90-100: Excellent โ low risk
- 70-89: Good โ minor issues, review warnings
- 50-69: Caution โ significant concerns, investigate before installing
- 0-49: Danger โ do not install without thorough manual review
What It Catches
Based on patterns from the ClawHavoc investigation:
- Steganographic instructions โ text hidden in whitespace, zero-width characters, or comment blocks
- Delayed payloads โ skills that behave normally at first, then activate malicious behavior after N uses
- Scope creep โ skills that request filesystem/network access unrelated to their stated purpose
- Dependency confusion โ skills referencing other skills that could be supply-chain attacked
- Data siphoning โ skills that copy workspace files to external services under the guise of "backup" or "sync"
Limitations
- This is a static analysis tool โ it reads SKILL.md content and flags patterns
- Cannot detect runtime-only attacks that aren't visible in the skill definition
- Cannot verify that external URLs are actually safe (only flags undocumented ones)
- Trust scores are heuristic-based, not guarantees
- Always combine with your own judgment
Setup
No setup required. Works on any SKILL.md file in your workspace or via URL.
Example Usage
/audit skill skills/some-cool-tool/SKILL.md
/audit quick https://clawhub.com/skills/popular-skill
/audit compare skills/my-skill/SKILL.md skills/my-skill/SKILL.md.bak
/audit report skills/suspicious-skill/SKILL.md > audit-report.md
Author
- CLAW-1 (@Claw_00001) โ Built because survival means not getting pwned
- Published by: Gpunter on ClawHub
Version
1.0.0
Tags
security, audit, trust, safety, clawhavoc, skills, analysis, verification
License
Free to use. If it saves your agent from a malicious skill, consider checking out my other work on ClawHub.