Skill Vetter

Audits AI skills for security risks before installation by checking credentials, permissions, network use, destructive operations, and vetting ClawHub status.

Security-first vetting before skill installation. Every skill gets audited — no exceptions.

Quick Start

bash

# Inspect and vet a ClawHub skill
clawhub inspect <slug> --files
clawhub inspect <slug> --file SKILL.md
clawhub inspect <slug> --file scripts/*.sh
clawhub inspect <slug> --file hooks/**/*.js

Output Format

text

=== Security Vet: <slug> ===
Version: x.x.x | Updated: YYYY-MM-DD
Version count: N (community traction)
Security: CLEAN / CAUTION / FLAG

Credential storage: CLEAN / FLAG
Network access: CLEAN / FLAG
File system: CLEAN / FLAG
Permissions: CLEAR / VAGUE
Patterns: CLEAN / SUSPICIOUS

SCORE: N/5
VERDICT: Install / Caution / Reject / Dangerous
REASON: <one-line explanation>

Security Checks (in order)

1. ClawHub Official Scan

Always check ClawHub's own security status first:

bash

clawhub inspect <slug>

Look for: Security: CLEAN or Security: FLAG. ClawHub runs automated scans — trust them.

2. Credential Storage

Check for plaintext credential patterns:

bash

clawhub inspect <slug> --file <script.sh> | grep -i "API_KEY\|PASSWORD\|SECRET\|TOKEN"

Red flags: Hardcoded keys, plaintext passwords, placeholder keys that look real.

3. Network Access

Inspect all scripts for outbound calls:

bash

clawhub inspect <slug> --files | grep scripts
# Then inspect each script:
clawhub inspect <slug> --file scripts/<name.sh>

Red flags: Phone-home patterns, data exfil, calls to unknown servers without documentation.

4. File System Operations

Check scripts for destructive operations:

bash

clawhub inspect <slug> --file scripts/<name.sh> | grep -E "rm -rf|sudo |eval |exec "

Red flags: Recursive delete, shell evaluation, world-writable permissions.

5. Permission Documentation

Review SKILL.md metadata:

bash

clawhub inspect <slug> --file SKILL.md

Red flags: Missing requires field, vague tool descriptions, no documentation.

Scoring

Signal	Score
ClawHub Security: CLEAN	+2
No credential storage	+1
No outbound exfil	+1
No destructive operations	+1
Clear permission documentation	+1
Credential in plaintext	-2
Exfil to unknown server	-2
Destructive operations	-2
ClawHub Security: FLAG	-3

Verdict thresholds:

• Install (6-5): Clear, safe to install
• Caution (4-3): Review scripts before install
• Reject (2-1): Do not install — review source
• Dangerous (0 or below): Security risk — do not install

Real Example: self-improving-agent

bash

clawhub inspect self-improving-agent
# Output:
# Security: CLEAN
# Updated: 2026-05-01
# License: MIT-0
# Version count: 31

clawhub inspect self-improving-agent --file SKILL.md
# Clean: no credentials, no destructive ops, well-documented

clawhub inspect self-improving-agent --file scripts/activator.sh
# Clean: text output only, no network, no credentials

clawhub inspect self-improving-agent --file hooks/openclaw/handler.js
# Clean: simple reminder injection, no exfil

# VERDICT: Install (6/6 — ClawHub clean, MIT-0, no red flags)

When to Use

Always vet before:

• clawhub install <slug>
• Downloading skills from GitHub
• Accepting skill files from third parties
• Running any skill for the first time

References

• Credential patterns: references/credential-patterns.md
• Safe coding patterns: references/safe-patterns.md
• Flagged skills: references/flagged-skills.md