← All articles
AI security

An AI security checklist for release reviews

Review access, data handling, and monitoring before putting an AI feature into production.

Review inputs, permissions, and actions

List where the feature accepts untrusted content, which data it can read, and which external actions it can trigger. Check that each action is authorized server-side. Use synthetic user roles to exercise allowed requests and ensure requests beyond their scope fail as expected.

Review what happens after a failure

Confirm that logs identify the operation and outcome without copying unnecessary personal data or secrets. Assign an owner to investigate repeated denials and define when a tool integration can be disabled. Test the recovery process before depending on it during an incident.

Review a feature before rollout

Imagine a support assistant that retrieves policy articles and can draft a case update. List its input sources, authorized readers, data returned to the model, and any action the system can execute. Specify which roles may draft or send a case update. Test each operation against synthetic records belonging to different accounts, and inspect what happens when a requested action is denied.

Check boundaries, retention, and monitoring

Verify that user identity and resource permissions are applied at the server. Confirm that write actions require the authorization and context appropriate to their consequences. Check that diagnostic events show the operation and outcome without collecting unnecessary content. Run a synthetic deletion and check whether the data disappears from expected locations. Record any components where loss or exposure would have an outsized impact.

Define stopping conditions before launch

List the failures that block rollout, such as a successful write outside a user’s scope or customer content appearing in a broadly accessible log. Separate them from issues that can be resolved through ordinary maintenance. Assign a decision owner and record evidence for each unresolved issue. Make sure reviewers can identify the deployed model, prompt, and tool configuration when comparing results.

Rehearse containment and recovery

Use a controlled exercise to check whether operators can disable a risky integration, restore a prior configuration, and verify the feature afterward. Document who has authority to make each change and where audit records are available. Prepare a short checklist for confirming that the system returns to expected behavior after a rollback or targeted fix.