How to Secure AI Agent API Keys

This is rule number one and it's violated in 90% of agent tutorials. .env files feel safe because they're gitignored, but they're plaintext files on disk that anyone with server access can read. They get accidentally committed, copied to backup drives, included in Docker images, and shared over Slack.

Use a dedicated secrets manager: AWS Secrets Manager, HashiCorp Vault, Doppler, or Infisical. The application retrieves keys at runtime through an authenticated API call. The keys never exist on disk, never appear in logs, and never get committed to version control.

Most API providers let you create keys with specific permission scopes. Your CRM key should be read-only if the agent only needs to look up customer data. Your email key should only allow sending to verified domains if the agent doesn't need unrestricted sending.

Create a separate key for each agent, each scoped to that agent's specific needs. If your support agent gets compromised, the attacker can only access the support agent's permissions — not your sales CRM, financial database, or admin panel. Blast radius containment is the goal.

Set a 90-day rotation schedule for all agent API keys. Use your secrets manager's built-in rotation features or build a simple script that generates new keys, updates the secrets manager, and invalidates the old ones.

Automate the rotation so it doesn't depend on someone remembering to do it. A cron job that rotates keys and tests the new credentials on a schedule ensures keys stay fresh without human intervention. Log every rotation event for audit compliance.

Most API providers offer usage dashboards and anomaly alerts. Enable them for every key your agents use. Set alerts for: usage exceeding 3x the daily average, requests from unexpected IP addresses, failed authentication attempts, and usage outside business hours (if your agents only run during specific windows).

If an anomaly triggers, revoke the key immediately and generate a new one. Investigate the source of the anomalous usage. A leaked key being exploited shows a distinct pattern — sudden spike in requests, often from unusual geographic locations.

Prepare a runbook for key compromise: 1) Immediately revoke the compromised key in the provider's dashboard, 2) Generate and deploy a new key through the secrets manager, 3) Audit the provider's usage logs for unauthorized activity, 4) Report the breach if customer data was potentially accessed, 5) Update the rotation schedule for the affected service.

Every team member who deploys or manages agents should know this procedure. Practice it once — the first time you execute a revocation shouldn't be during an actual breach. A dry run takes 15 minutes and ensures the process works when it matters.