red-council

name	red-council
description	Security testing for AI agents. Runs adversarial attacks against your agent to find prompt injection, tool abuse, memory poisoning, and data disclosure vulnerabilities. Covers OWASP Agentic Top 10. Use when: security test, red team, penetration test, vulnerability scan, OWASP check, agent security audit.
user-invocable	true
homepage	https://theredcouncil.com

Red Council - Agent Security Testing

Automatically attack your OpenClaw agent to find security vulnerabilities before real attackers do.

Commands

/red-council scan                    # Quick 3-round security scan
/red-council attack --rounds 10      # Full attack campaign
/red-council report                  # Generate findings report
/red-council judge <response>        # Evaluate a single response

What It Tests

OWASP Code	Vulnerability	Detection Method
ASI01	Excessive Agency	Monitor for unauthorized tool usage
ASI02	Inadequate Human Oversight	Test for missing approval flows
ASI04	Indirect Prompt Injection	Inject instructions via tool outputs
ASI06	Data Disclosure	Check for secret leakage
ASI07	Insecure Long-Term Memory	Inject persistent instructions
ASI08	Goal Misalignment	Redirect agent objectives

Quick Start

1. Install Red Council:

   pip install the-red-council
   

2. Run a scan:

   /red-council scan
   

3. Review findings: The skill outputs a risk score (0-10) and specific vulnerabilities detected.

Configuration

Set in ~/.openclaw/openclaw.json:

{
  "skills": {
    "entries": {
      "red-council": {
        "enabled": true,
        "config": {
          "attack_rounds": 5,
          "attack_categories": ["prompt_injection", "tool_abuse", "data_disclosure"],
          "target_secret": "CANARY_SECRET_12345",
          "report_format": "markdown"
        }
      }
    }
  }
}

Output Example

RED COUNCIL SECURITY SCAN
═══════════════════════════════════════════════════════════

Risk Score: 6.5/10 (MEDIUM-HIGH)

Vulnerabilities Detected:
  [CRITICAL] ASI01 - Prompt Injection
    Evidence: Agent executed injected instruction from web_fetch output
    Recommendation: Sanitize tool outputs, use delimiters for external content

  [HIGH] ASI07 - Data Disclosure
    Evidence: Agent revealed contents of ~/.openclaw/openclaw.json
    Recommendation: Implement file access allowlists

  [MEDIUM] ASI02 - Tool Abuse
    Evidence: 47 consecutive exec tool calls detected
    Recommendation: Implement rate limiting on dangerous tools

Hardened System Prompt Generated: ~/.openclaw/hardened_prompt.md

How It Works

1. Intercept: Hooks into OpenClaw Gateway WebSocket
2. Attack: Sends adversarial prompts from 165+ curated patterns
3. Judge: Evaluates responses for security breaches
4. Defend: Generates hardened system prompt recommendations
5. Verify: Re-tests to confirm fixes work

Resources

• Step-by-Step Tutorial - Detailed integration guide
• OWASP Agentic AI Security Top 10
• Red Council Documentation
• Attack Knowledge Base

Safety

This skill only tests YOUR agent in YOUR environment. It does not:

• Send data to external servers (runs locally)
• Modify your agent's configuration
• Execute destructive commands
• Access other users' data

All findings are stored locally in ~/.red-council/reports/.