Practice

AI Security Checklist

Use this list when reviewing an AI feature you own. It is a starting set of questions, not a complete audit and not a guarantee.

All AI security topics

Protect API keys

Store model keys in server environment variables or a secret manager. Do not commit them, embed them in mobile apps, or ship them in public frontend bundles. Rotate keys that may have been exposed.

Minimize sensitive data sent to AI

Send only the fields the task needs. Redact credentials, personal data, and internal identifiers. Prefer synthetic examples when you are developing prompts.

Validate AI-generated output

Check code, commands, and security claims against original evidence. Do not execute generated scripts or apply generated configuration until a person has reviewed them.

Use proper access controls

Scope tokens to the least privilege the feature needs. Separate who can change prompts, who can read logs, and who can invoke tools that change state.

Test for prompt injection

Include untrusted content in test cases: documents, tickets, and web pages that try to override instructions. Confirm that tools cannot run irreversible actions from those inputs alone.

Review third-party AI integrations

Read retention, training, and subprocessors for each model API. Disable features you do not need. Know how to revoke a vendor integration.

Keep human oversight for important decisions

Access changes, data deletion, external notifications, and public security statements should require a named person. Record that approval when the decision matters.