An exposed AI key can turn into a six-figure day

Google Threat Intelligence Group and Mandiant describe multiple 2026 cases in which attackers used enterprise access to large language models without permission. In individual investigations, fraudulent use exceeded $100,000 in a day, with cumulative charges at some organizations reaching $1 million before access was revoked. Those are observed cases, not a prediction of what every stolen key will cost.

Investigators found several entry routes: exposed model API keys in public repositories, compromised cloud identities and service accounts, and access to internal code and secret stores. Some attackers tested keys with tiny requests before scaling usage or routing it through resale infrastructure.

1. 1

   Inventory credentials and owners. Find model API keys, gateway virtual keys and service-account permissions. Remove unused access and keep keys out of repositories and build output.

2. 2

   Set usage and cost signals. Alert on unusual request volume, model changes, region changes and daily spend. Test whether the right team receives the alert quickly enough to revoke access.

3. 3

   Rotate and investigate. If misuse is suspected, disable the exposed key and inspect identity, gateway, repository and billing logs. Check for additional credentials or privileged accounts created during the intrusion.

AI usage often inherits the same identity and secret-management weaknesses as the rest of a cloud account. Assign a real owner to each model credential, limit its permissions and budget, and make revocation a practiced step. A cost spike should trigger an access review, not only a billing dispute.