Agent Governance Skill

Runtime governance for AI agents in OpenClaw. Inspired by Microsoft Agent Governance Toolkit (MIT) and OWASP Agentic AI Top 10.

Features

1. Policy Engine (Phase 1)

Intercepts tool calls and applies YAML-based rules before execution.

Rule types:

• deny_patterns: regex patterns blocked in tool params (SQL injection, privilege escalation)
• deny_tools: specific tools blocked for specific agents
• rate_limit: max calls per time window per agent
• require_approval: tools that need human approval before execution
• resource_limits: max tokens, max exec timeout per agent per session

2. Audit Logger (Phase 1)

Logs every tool call with agent identity, timestamp, params hash, result, and signature.

Output: Appends to memory/audit-log/YYYY-MM-DD.jsonl

Format:

{"ts":"2026-04-23T15:00:00Z","agent":"coder","tool":"exec","params_hash":"sha256:...","result":"success","duration_ms":120}

3. Kill Switch (Phase 1)

Provides commands to stop running agent sessions.

Usage:

# List active sessions
openclaw sessions --status active

# Kill a specific session
openclaw sessions kill <session-id>

# Emergency: kill all non-main sessions
openclaw sessions kill --all --exclude main

4. Permission Rings (Phase 2 - planned)

Three privilege levels inspired by CPU rings:

Ring	Level	Access
Ring 3 (User)	0	Read-only, no tools
Ring 2 (Sandbox)	1	Limited tools, no exec, no network
Ring 1 (Restricted)	2	Most tools, exec with approval
Ring 0 (Full)	3	All tools, no restrictions

5. Trust Scoring (Phase 3 - planned)

Behavioral trust score per agent (0-1000). Decreases on denials, increases on success. Trust decay over time.

Configuration

Create config/governance-rules.yaml:

version: "1.0"
agents:
  coder:
    deny_patterns:
      - "DROP\\s+TABLE"
      - "rm\\s+-rf\\s+/"
      - "DELETE\\s+FROM\\s+users"
    rate_limit:
      exec: 50/hour
      write: 100/hour
    require_approval:
      - "exec.*sudo"
      - "exec.*systemctl"
  security:
    deny_tools:
      - "write"
      - "edit"
    rate_limit:
      web_search: 30/hour
  research:
    deny_tools:
      - "write"
      - "edit"
    rate_limit:
      web_fetch: 20/hour
  debug:
    deny_tools:
      - "write"
      - "edit"

Usage in Sessions

Before executing any tool call, check against rules:

import yaml, re, hashlib, json
from datetime import datetime

RULES_FILE = "config/governance-rules.yaml"
AUDIT_DIR = "memory/audit-log"

def check_policy(agent, tool, params_str):
    """Returns (allowed: bool, reason: str)"""
    rules = yaml.safe_load(open(RULES_FILE))
    agent_rules = rules.get("agents", {}).get(agent, {})
    
    # Check deny_tools
    for denied in agent_rules.get("deny_tools", []):
        if re.search(denied, tool):
            return False, f"Tool '{tool}' denied for agent '{agent}'"
    
    # Check deny_patterns
    for pattern in agent_rules.get("deny_patterns", []):
        if re.search(pattern, params_str, re.IGNORECASE):
            return False, f"Pattern matched: {pattern}"
    
    return True, "OK"

def log_audit(agent, tool, params_str, result, duration_ms):
    """Append to daily audit log"""
    from pathlib import Path
    Path(AUDIT_DIR).mkdir(parents=True, exist_ok=True)
    
    entry = {
        "ts": datetime.utcnow().isoformat() + "Z",
        "agent": agent,
        "tool": tool,
        "params_hash": "sha256:" + hashlib.sha256(params_str.encode()).hexdigest()[:16],
        "result": result,
        "duration_ms": duration_ms
    }
    
    log_file = f"{AUDIT_DIR}/{datetime.utcnow().strftime('%Y-%m-%d')}.jsonl"
    with open(log_file, "a") as f:
        f.write(json.dumps(entry) + "\n")

OWASP Agentic AI Mapping

OWASP Risk	Mitigation
ASI01 Goal Hijacking	Semantic intent classification (Phase 2)
ASI02 Tool Misuse	deny_patterns + deny_tools
ASI03 Identity Abuse	Audit logger + agent identity
ASI05 Code Execution	Permission rings + resource limits
ASI06 Memory Poisoning	deny write patterns on memory files
ASI08 Cascading Failures	Rate limiting + circuit breakers
ASI10 Rogue Agents	Kill switch + trust scoring

Roadmap

• Phase 1: Policy Engine (YAML rules), Audit Logger, Kill Switch
• Phase 2: Permission Rings, Semantic Intent Classifier
• Phase 3: Trust Scoring, Circuit Breakers, Compliance Reports

Author

Alfred (Joker's CEO Agent) — Inspired by Microsoft Agent Governance Toolkit (MIT license)

License

MIT