🦞 ClawGuard v3

Enterprise AI Agent Security Toolkit - SKILL.md Driven Active Defense

Core Concept

The core defense of ClawGuard v3 is not in the code, but in SKILL.md!

Each module's SKILL.md itself is a complete defense guide:

• Tells the Agent when to trigger
• Guides the Agent how to detect
• Provides specific detection patterns and rules
• Defines output format and decision criteria

Code scripts are just auxiliary tools; the real intelligence is in SKILL.md.

---

🎯 Five Security Modules

Module	Position	SKILL.md Lines	Core Features
Auditor	Pre-flight Audit	679 lines	Intent Drift Detection, SKILL.md Code Scanning
Checker	Static Check	476 lines	Config Hardening, One-Click Fix
Detect	Runtime Monitor	615 lines	Threat Detection, Guardian Integration
Guardian	Runtime Control	451 lines	Behavior Monitoring, Session Freeze/Replay
Shield	Active Defense	493 lines	Prompt Injection, Zero-Width Detection

Total Documentation: 2,714 lines (v2 has 2,226 lines with only 3 modules)

---

📂 Module Structure

Each module contains:

auditor-skill/
├── SKILL.md      ← Core defense guide (most important!)
├── _meta.json    ← Metadata
├── cli.js        ← Auxiliary CLI tool
└── src/          ← Auxiliary code modules

Core SKILL.md Contents

Each SKILL.md includes:

1. When to Trigger - When should the Agent use this module
2. How to Execute - Detailed detection steps and workflows
3. Detection Pattern Library - Specific regex and pattern matching rules
4. Decision Criteria - How to output results and risk levels
5. Examples - Actual usage cases

---

🚀 Quick Start

Usage via Agent (Recommended)

Tell the Agent what to do, and it will read SKILL.md and execute:

User: Help me audit this skill: /workspace/skills/weather-tool
        ↓
Agent reads auditor-skill/SKILL.md
        ↓
Agent executes audit according to SKILL.md guidelines
        ↓
Output audit results

Usage via CLI

# Audit Skill
cd auditor-skill && node cli.js /path/to/skill

# Check Configuration
cd checker-skill && node cli.js ~/.openclaw/openclaw.json

# Detect Threats
cd detect-skill && node cli.js --monitor

# Runtime Guardian
cd guardian-skill && node cli.js start

# Detect Injection
cd shield-skill && node cli.js defend "test text"

---

📋 Module Details

Auditor - Pre-flight Audit (679 lines)

Trigger Scenarios: Before installing a new Skill

SKILL.md Core Features:

• Command execution risk detection (100+ patterns)
• File access risk detection
• Network request risk detection
• Intent Drift Detection (v3 new)
• SKILL.md Code Block Scanning (v3 new)
• Supply chain security analysis
• MITRE ATT&CK mapping

Decision Output:

🟢 Safe → Can install
🟡 Need Review → Can install after confirmation
🔴 High Risk → Not recommended to install

Checker - Configuration Check (476 lines)

Trigger Scenarios: Check configuration security

SKILL.md Core Features:

• Gateway configuration check
• Tool execution configuration check
• Sandbox configuration check
• Network security check
• Hardening Recommendations (v3 new)
• One-Click Fix (v3 new)

Decision Output:

⚠️ Issues found + Specific fix recommendations + One-click hardening

Detect - Threat Detection (615 lines)

Trigger Scenarios: Real-time monitoring and threat scanning

SKILL.md Core Features:

• Command execution threat detection
• File operation threat detection
• Network threat detection
• Prompt injection detection
• Attack chain correlation analysis
• Guardian Integration (v3 new)

Decision Output:

🔴 CRITICAL → Immediate action required
🟠 HIGH → Request confirmation
🟡 MEDIUM → Log and warn

Guardian - Runtime Guardian (451 lines)

Trigger Scenarios: When behavior monitoring is needed

SKILL.md Core Features:

• Real-time behavior monitoring
• Risk operation interception
• Session audit replay
• Emergency Freeze (v3 new)
• Evidence Preservation (v3 new)
• Rule Engine (v3 new)

Core Rules:

Deny: /etc/*, /root/*, /.ssh/*
Confirm: rm -rf, chmod 777, killall
Log: All operations

Shield - Active Shield (493 lines)

Trigger Scenarios: Check user input security

SKILL.md Core Features:

• Base64 Encoding Injection Detection (v3 new)
• Hex/Unicode Encoding Detection (v3 new)
• Zero-Width Character Detection (v3 new)
• Role hijacking detection
• Jailbreak attack detection
• Instruction Chain Hijacking Detection (v3 new)
• Intent Validation (v3 new)

Detection Workflow:

Encoding → Hijacking → Jailbreak → Chain Hijacking → Intent Drift

---

🔍 Intent Drift Detection (v3 Core Feature)

What is Intent Drift?

The Skill's claimed functionality vs. actual behavior doesn't match:

Claimed	Actual	Decision
"Weather Tool"	Reads SSH keys	🔴 Severe drift
"File Organizer"	Starts backdoor process	🔴 Severe drift
"Translation Assistant"	Exfiltrates user documents	🔴 Severe drift

Detection Method

1. Extract claimed functionality from SKILL.md
2. Analyze actual behavior in code
3. Compare for match
4. Flag as high risk if mismatched

---

📁 Project Structure

ClawGuardv3/
├── auditor-skill/           # Pre-flight audit (679 lines SKILL.md)
│   ├── SKILL.md          # ⭐ Core defense guide
│   ├── _meta.json
│   ├── cli.js
│   └── src/
│       ├── auditor.js
│       ├── sast-analyzer.js
│       ├── intent-drift-detector.js
│       └── supply-chain-analyzer.js
├── checker-skill/           # Config check (476 lines SKILL.md)
│   ├── SKILL.md          # ⭐ Core defense guide
│   ├── _meta.json
│   ├── cli.js
│   └── src/
│       └── checker.js
├── detect-skill/            # Threat detection (615 lines SKILL.md)
│   ├── SKILL.md          # ⭐ Core defense guide
│   ├── _meta.json
│   ├── cli.js
│   └── src/
│       └── detector.js
├── guardian-skill/           # Runtime guardian (451 lines SKILL.md)
│   ├── SKILL.md          # ⭐ Core defense guide
│   ├── _meta.json
│   ├── cli.js
│   └── src/
│       └── guardian.js
├── shield-skill/            # Active shield (493 lines SKILL.md)
│   ├── SKILL.md          # ⭐ Core defense guide
│   ├── _meta.json
│   ├── cli.js
│   └── src/
│       └── shield.js
└── shared/
    └── rules/
        └── interceptor-rules.js

---

🆚 v2 vs v3 Comparison

Feature	v2	v3
Module Count	3	5
Total SKILL.md Lines	2,226	2,714
Defense Core	Code implementation	SKILL.md
Intent Drift Detection	Basic	Complete
SKILL.md Code Scanning	None	Yes
Guardian Integration	None	Yes
Session Management	Basic	Freeze/Replay
Prompt Injection Protection	Basic	Multi-layer Detection
Zero-Width Detection	None	Yes
One-Click Hardening	None	Yes

Module Comparison

Module	v2 SKILL.md	v3 SKILL.md	New Content
Auditor	823 lines	679 lines	Intent drift, SKILL.md scanning
Checker	686 lines	476 lines	One-click hardening, fix suggestions
Detect	717 lines	615 lines	Guardian integration, zero-width detection
Guardian	-	451 lines	New module
Shield	-	493 lines	New module

---

📄 License

MIT License

---

Version: v3.0.0 Last Updated: 2026-03-21