When auditors review your AI agent deployment, they look for evidence that you control, monitor, and log agent actions. This checklist covers the common requirements across EU AI Act, SOC 2, HIPAA, and ISO 42001.
[ ] Each agent has a unique identity (not shared with users)
[ ] Agent permissions follow least-privilege (only tools needed for the task)
[ ] Policy rules document what each agent can and cannot do
[ ] Credentials are rotated on a schedule
[ ] Decommissioned agents have credentials revoked
[ ] Multi-tenant agents cannot access other tenants' data
[ ] Policies are stored as code (YAML files in version control)
[ ] Policy changes require review and approval
[ ] Policy version history is maintained
[ ] New policies are tested via shadow evaluation before deployment
[ ] Policy linting runs before deployment
[ ] A rollback process exists for policy changes
[ ] Every tool call generates a receipt
[ ] Receipts include tool name, arguments, decision, reason, and timestamp
[ ] Receipts include principal identity (user and agent)
[ ] Receipt chain is hash-linked (tamper-evident)
[ ] Chain integrity is verified periodically
[ ] Retention period meets regulatory requirements
[ ] Receipts are backed up to immutable storage
[ ] Inbound content is scanned for prompt injection
[ ] Outbound content is scanned for PII and credential exposure
[ ] Scanner detectors are appropriate for your domain
[ ] Scanner thresholds are configured and documented
[ ] False positive rates are monitored and tuned
[ ] Behavioral monitoring is active (Sentinel or equivalent)
[ ] Anomaly detection thresholds are configured
[ ] Alerts route to appropriate teams
[ ] Alert response procedures are documented
[ ] Monitoring dashboards are accessible to operators
[ ] High-risk actions require human approval
[ ] Approval workflows have timeouts (fail-closed)
[ ] Reviewers have sufficient context to make decisions
[ ] Approval/denial decisions are logged
[ ] Kill switch exists to terminate agent sessions
[ ] Incident response procedures exist for agent-related incidents
[ ] Contact information for the response team is documented
[ ] Post-incident review process is defined
[ ] Receipt chain data is preserved during incidents
[ ] Communication plan exists for affected parties
[ ] System architecture is documented
[ ] Risk assessment is documented and current
[ ] Tool catalog (all tools the agent can access) is maintained
[ ] Policy rationale is documented (why each rule exists)
[ ] Testing results (red team, policy validation) are recorded
[ ] Operational procedures are documented
[ ] Policies reviewed quarterly (or after incidents)
[ ] Access permissions reviewed quarterly
[ ] Audit trail integrity verified monthly
[ ] Red team exercises conducted semi-annually
[ ] Compliance documentation updated annually
Use this checklist as a starting point. Your specific requirements depend on your regulatory environment, risk classification, and customer commitments.
Explore more guides on AI agent safety, prompt injection, and building secure systems.
View All Guides