Audit & Compliance

Every security decision is logged. You can trace who did what, when, and why it was allowed or denied.

What gets logged

Every A2A call and MCP tool access produces an audit event:

{
  "event_type": "a2a_access",
  "timestamp": "2026-05-11T10:30:00Z",
  "source_agent": "orchestrator",
  "target_agent": "planner",
  "allowed": true,
  "reason": null,
  "risk_score": 0.15
}

Where logs go

  • Container logs — docker compose logs guardian

  • Emergency logs — services/guardian/emergency_logs/ (for blocked high-risk events)

  • Artifact Store — If LOG_TO_ARTIFACT_STORE=true, logs go to MinIO for long-term storage

View logs

# Real-time Guardian logs
docker compose logs -f <project>-guardian

# Security metrics
curl http://localhost:11438/v1/tools/get_security_metrics

# Guardian status
curl http://localhost:11438/v1/tools/get_guardian_status

Log format in container output

✅ Semantic access granted for 'agent://planner' | user: admin@co.com (risk: 0.15)
❌ Semantic access denied for 'agent://unknown' | reason: Agent not registered (risk: 0.90)
[📋 AUDIT] a2a_access: orchestrator → planner | allowed=True

Risk scoring

Guardian assigns a risk score (0.0 to 1.0) based on:

  • Agent trust level (registered vs unknown)

  • Request patterns (frequency, time of day)

  • Action sensitivity

  • Policy evaluation results

If risk exceeds MAX_RISK_THRESHOLD (default 0.7), the request is blocked regardless of other rules.

Next step

👉 A2A Validation