The report
Fireworks AI's security bulletin, last updated October 10, describes unauthorized access to environment variables used in customer evaluation jobs. The company identified the activity on October 6, notified a limited number of affected accounts and extended its credential rotations on October 10. The exposure is specific to the evaluation environment, not a claim that inference prompts or customer models were accessed. Fireworks' incident dossier preserves the supported scope.
What changed in the latest bulletin
Fireworks says internal credentials were used to read evaluation-job variables, and it believes stored API keys and tokens were captured in some affected accounts. It began containment on October 6 and customer notifications on October 7. Employee API keys were rotated October 9; a completed credential review led to additional Fireworks user API-key revocations and service-account token rotations on October 10. The dated response sequence matters because the first public bulletin was not the last action.
The company says its investigation has not detected unauthorized access to customer models, datasets, training data, inference prompts and outputs, or personal information within Fireworks' own security perimeter. That is a qualified investigative finding, not evidence that every downstream service reached by a possibly exposed third-party credential remained untouched. Fireworks says the investigation and review of calls made with exposed credentials are continuing. Official incident bulletin.
Whose credentials need to change
A Fireworks-issued key can be revoked by Fireworks. A token for an external provider must be revoked or rotated at its issuing service; deleting it from the evaluation secret store does not invalidate it. Fireworks has given identified affected customers direct notices and recommends checking the access logs and resources reachable through those credentials. Issuer-side rotation requirement.
The entry point into internal credentials, final affected-account count and any external use of customer secrets remain undisclosed. The concrete boundary problem is that a provider-held identity could reach customer-supplied evaluation secrets, irrespective of zero-data-retention controls for ordinary inference traffic.