pincer（技能安全扫描）

pincer 🛡️

Security-first wrapper for clawhub install. Scans skills for malware, prompt injection, and suspicious patterns before installation.

Why?

Agent skills are powerful — they're basically executable documentation. The ClawHub ecosystem has already seen malware campaigns distributing infostealers via innocent-looking skills. pincer adds a security layer before you install anything.

Install

# From ClawHub
clawhub install pincer

# Or manually
chmod +x ./scripts/pincer.sh
ln -sf "$(pwd)/scripts/pincer.sh" ~/.local/bin/pincer

Dependencies:

• clawhub — for fetching skills
• uvx — for mcp-scan (brew install uv)
• jq — for JSON parsing

Usage

Safe Install

# Instead of: clawhub install some-skill
pincer install some-skill

# With specific version
pincer install some-skill@1.2.0

Scan Without Installing

# Scan a ClawHub skill
pincer scan some-skill

# Scan a local directory
pincer scan ./path/to/skill

# JSON output for automation
pincer scan some-skill --json

Audit Installed Skills

# Quick-scan all installed skills
pincer audit

# JSON output
pincer audit --json

Manage Trust

# Add trusted publisher (auto-approve clean skills)
pincer trust add steipete

# Remove from trusted
pincer trust remove old-publisher

# Block a publisher or skill
pincer trust block suspicious-dev
pincer trust block malware-skill

# Unblock
pincer trust unblock redeemed-dev

# List all trust settings
pincer trust list

View History

# See what you've installed
pincer history

# JSON output
pincer history --json

Configuration

# Show current config
pincer config show

# Edit in $EDITOR
pincer config edit

# Reset to defaults
pincer config reset

What It Checks

Via mcp-scan (Invariant Labs)

• Prompt injection attacks
• Malware payloads in natural language
• Tool poisoning
• Sensitive data exposure
• Hard-coded secrets

Additional Pattern Detection

Publisher & Provenance

• Publisher reputation (trusted list)
• Download count threshold
• Skill age threshold
• Blocklist checking

Binary Detection

• Scans for bundled executables
• Flags Mach-O, ELF, PE32 binaries

Risk Levels

Configuration

Config: ~/.config/pincer/config.json

{
  "trustedPublishers": ["openclaw", "steipete", "invariantlabs-ai"],
  "blockedPublishers": [],
  "blockedSkills": [],
  "autoApprove": "clean",
  "logInstalls": true,
  "minDownloads": 0,
  "minAgeDays": 0
}

Examples

Clean Install

$ pincer install bird
🛡️ pincer v1.0.0

  → Fetching bird from ClawHub...
  Publisher: steipete (trusted)
  Stats: 7363 downloads · 27 ★ · created 1 month ago

🛡️ pincer Scanning bird...

  → Running mcp-scan...
  ✅ mcp-scan: passed
  → Checking for suspicious patterns...
  ✅ Pattern check: passed
  → Checking external URLs...
  ✅ URL check: passed
  → Checking for bundled binaries...
  ✅ Binary check: passed

Risk Assessment:
  ✅ CLEAN — No issues detected

  → Auto-approved (clean + trusted config).
  → Installing bird...
  ✅ Installed successfully!

Dangerous Skill Blocked

$ pincer install sketchy-tool
🛡️ pincer v1.0.0

  → Fetching sketchy-tool from ClawHub...
  Publisher: newaccount (unknown)
  Stats: 12 downloads · 0 ★ · created 2 days ago

🛡️ pincer Scanning sketchy-tool...

  → Running mcp-scan...
  🚨 mcp-scan: high-risk warnings
  → Checking for suspicious patterns...
  🚨 Pattern check: suspicious patterns found
    • curl/wget piped to shell
    • macOS quarantine removal (xattr)
  → Checking external URLs...
  ⚠️ URL check: external URLs found
    • http://sketchy-domain.xyz/install
  → Checking for bundled binaries...
  ✅ Binary check: passed

Risk Assessment:
  🚨 DANGER — Suspicious patterns detected
    • mcp-scan: high-risk patterns detected
    • curl/wget piped to shell
    • macOS quarantine removal (xattr)

  ☠️ Install blocked. Use --force to override (not recommended).

Credits

• mcp-scan by Invariant Labs — core security scanning
• 1Password Security Research — threat analysis that inspired this tool
• Snyk ToxicSkills Report — ecosystem threat research

License

MIT

Stay safe out there. 🛡️