Two stolen AI keys show why model quotas need abuse controls

Wiz traced two compromised enterprise keys from validation calls to large Bedrock traffic bursts. Teams need fast credential revocation and usage alarms before model access becomes a resale channel.
Wiz Research reports that two compromised enterprise AI credentials were tested within 30 minutes of reported vendor credential access, then used for heavy Amazon Bedrock traffic days later. One key generated about 800,000 calls over 30 hours; the other generated about 1.7 million calls over 13 hours. The practical issue is the gap between the first test call and revocation.
Wiz also counts at least nine AI developer platforms that disclosed incidents during October. It says public reports do not establish a single vulnerability or actor linking them. The two credential cases show downstream abuse after vendor breaches, but Wiz does not definitively connect the resellers it identified to the original compromises.
A test call can be the first billing signal
In both cases, Wiz observed an initial InvokeModel request from a residential address, followed by a low volume pattern of checks that kept the credentials warm. The larger traffic came later, after model and quota probing. That sequence means an account can look quiet after a credential exposure while an attacker confirms which access is saleable.
The two keys had different policy scopes. The first permitted global cross-region inference profiles, and Wiz says about 90% of its 800,000 calls succeeded. The second was limited to US profiles; many initial requests for global profiles failed, yet an actor found working model IDs and later made over a million successful calls. Narrower scope reduced some calls but did not end the abuse.
The first unauthorised model call is a revocation deadline, not a harmless probe.
Spend controls need an owner
Wiz estimates the two cases likely exceeded $100,000 in usage at list price. That estimate is not a verified invoice, but it shows why security and finance teams need a shared response path. A spike in InvokeModel volume, model catalogue probing or calls from new networks should have a named owner who can disable the credential quickly.
The useful detection signal is a change in behaviour, not a fixed list of attacker addresses. Compare model calls with each workload's usual source networks, regions, models and hourly volume. Alert on quota enumeration and sudden switches to premium models. Keep an approved emergency procedure for revoking vendor-held keys, then issue replacements with the smallest viable model and region scope.
Treat vendor access as a live boundary
Teams should inventory which developer platforms hold long lived cloud or model credentials, whether those credentials can be isolated per vendor, and who receives breach notices. Where possible, use short lived identities and separate billing or workload boundaries so one compromised integration cannot spend across an entire account.
After any platform incident, check the first unauthorised use rather than waiting for a bill. Wiz's cases show that a short validation call can precede a long quiet period and then a rapid flood. The control worth testing is how quickly the organisation can detect, revoke and account for that access.
Sources
- “Black October” for AI Developer Platforms · Wiz Research
Written by the Devence Lab research team.




