AI Agent Compliance Checklist

This checklist covers the technical controls needed for AI agent compliance. It applies whether you are targeting EU AI Act, SOC 2, or simply building a responsible agent deployment.

Authorization Controls

• [ ] Policy engine deployed with fail-closed default (no policy = deny)
• [ ] Every tool call evaluated against policy before execution
• [ ] Role-based access controls for tool access
• [ ] Parameter-level constraints on tool calls
• [ ] Rate limits per tool, per agent, and per user
• [ ] Budget caps on cumulative resource usage
• [ ] No path from agent to tool execution that bypasses the policy engine

Content Safety

• [ ] Input scanning for prompt injection on all user input
• [ ] Tool description scanning for injection on MCP server connections
• [ ] Tool response scanning before re-entry into model context
• [ ] Output scanning before delivery to user
• [ ] Encoding attack detection (base64, hex, unicode)
• [ ] Multi-language detection coverage

Human Oversight

• [ ] Approval workflows for high-risk actions
• [ ] Kill switch that operates independently of the agent
• [ ] Kill switch tested and verified (not just implemented)
• [ ] Monitoring dashboard for real-time agent observation
• [ ] Escalation procedures defined for different incident severities
• [ ] Reviewers trained and roles documented

Audit Trail

• [ ] Every tool call produces an audit record
• [ ] Every policy evaluation produces an audit record
• [ ] Every approval decision produces an audit record
• [ ] Records include timestamp, identity, action, parameters, decision, outcome
• [ ] Records are hash-chained for tamper evidence
• [ ] Storage is append-only (no UPDATE or DELETE capability)
• [ ] Retention period meets regulatory requirements
• [ ] Chain verification runs periodically
• [ ] Export capability for audit review

Monitoring

• [ ] Behavioral monitoring active on tool-call patterns
• [ ] Anomaly detection algorithms configured with appropriate thresholds
• [ ] Alerting tiers defined (log, notify, escalate, automate)
• [ ] Baseline established from normal operation
• [ ] Monitoring data retained for trend analysis

Testing

• [ ] Red team testing conducted with current payload corpus
• [ ] Detection rates measured per attack category
• [ ] Authorization bypass testing completed
• [ ] Kill switch tested under load
• [ ] Approval workflow tested end-to-end
• [ ] Regression testing in CI/CD pipeline
• [ ] Test results documented and retained

Documentation

• [ ] Risk assessment completed and documented
• [ ] System architecture documented including safety controls
• [ ] Tool access inventory maintained
• [ ] Policy rules documented with rationale
• [ ] Incident response playbook written and accessible
• [ ] Oversight roles and responsibilities documented
• [ ] Known limitations documented

Operational Procedures

• [ ] Incident response plan defined and tested
• [ ] On-call rotation for agent monitoring
• [ ] Regular red team exercises scheduled
• [ ] Policy review cadence established
• [ ] Monitoring threshold review cadence established
• [ ] Audit trail integrity review cadence established

Not every item applies to every deployment. Low-risk agents in internal tools can skip some items. High-risk agents in regulated domains need all of them. Use the EU AI Act classification to determine your scope.