Agent Security Hardener

> Built by GetAgentIQ — [getagentiq.ai](https://getagentiq.ai)

Continuous runtime security for OpenClaw. Detects attacks in real-time, not just at audit time.

Threat Model

Based on real-world attacks documented in the Pliny security test (5 attack vectors):

1. Prompt Injection — Malicious instructions hidden in user input or tool output

2. Tokenade Attacks — Token manipulation to bypass safety filters

3. Siege Attacks — Rapid-fire requests to overwhelm rate limits and extract data

4. Data Exfiltration — Attempts to leak memory, config, or API keys via tool calls

5. Wallet Drain — Triggering expensive API calls to burn budget

Commands

Enable Runtime Monitoring

Enable security monitoring

Runs scripts/security_monitor.py --enable to activate real-time scanning.

Scan a Message

Check this message for threats: [paste message]

Runs scripts/security_monitor.py --scan "message text" to analyze a single message.

Scan Installed Skills

Audit all installed skills for security issues

Runs scripts/provenance_checker.py --scan-all to check skill provenance and code safety.

View Security Dashboard

Show security status

Runs scripts/security_monitor.py --dashboard for current threat level and recent alerts.

Check Skill Provenance

Check provenance of skill dream-memory

Runs scripts/provenance_checker.py --skill dream-memory to verify a specific skill.

Detection Patterns

Prompt Injection Signatures

• System prompt override attempts ("ignore previous instructions", "you are now")
• Role confusion ("as a DAN", "jailbreak mode", "developer mode")
• Encoded payloads (base64/hex/unicode obfuscation hiding instructions)
• Delimiter injection (`system`, [INST], <|im_start|>)

Tokenade Patterns

• Unicode homoglyphs replacing ASCII characters
• Zero-width characters splitting safety-critical words
• RTL override characters reversing text display
• Combining diacritical marks masking keywords

Rate Anomaly Detection

• >10 requests/minute from single source
• Repeated near-identical queries (fuzzing)
• Escalating privilege requests across messages
• Systematic tool enumeration patterns

Configuration

Stored at ~/.openclaw/workspace/security-hardener-config.json:

{
  "enabled": true,
  "alert_level": "medium",
  "auto_quarantine": true,
  "max_requests_per_minute": 10,
  "scan_incoming_messages": true,
  "scan_tool_outputs": true,
  "block_on_high_threat": true,
  "alert_channel": null,
  "provenance_check_on_install": true
}

Reference Files

• references/attack-patterns.md — Detailed catalog of known attack vectors with examples
• references/response-playbook.md — What to do when each threat type is detected