An exposed credential gave an unauthorized party access to a private Hugging Face repository containing prompt samples from Together AI’s serverless inference service. Together AI says the samples were downloaded after the credential appeared in a public code repository during development of its public GitHub project, xorl.
How the access happened
Together AI says it collected the samples as part of routine software testing, with authorization for that use. The samples came from requests to three models: Minimax M3, GLM 5.2 and Kimi 2.6.
An AI coding agent committed a credential that provided access to the private repository while developers were working on xorl. Together AI’s automated scanning detected the exposed credential, and the company reverted the commit. During the exposure window, however, an unauthorized party obtained and used the credential to download the stored samples.
What the samples contained
The affected requests were made primarily on June 24, July 4, July 8 and July 14, 2026. Together AI has not said how many prompt samples were downloaded or what each one contained.
The party searched the samples for credentials to other services and tested whether those credentials worked. Together AI says five of the credentials tested were canary credentials associated with a Dragos security experiment; it does not report whether any other credentials were valid.
What the investigation has found
Together AI says it hired an external forensic investigation firm and worked with law enforcement. Its review has found no evidence of prompt leakage beyond this incident, but the company is still examining the affected prompts for sensitive personal information or other customer data.
The company says it has no reason to believe the unauthorized party leaked the data further and understands that the samples have now been deleted. It says it will contact affected customers directly as needed, with findings relevant to them.
Changes to accounts and development
Together AI says it has strengthened controls around development workflows, credentials and customer data. The changes include using scoped or short-lived credentials, and keyless access where supported, to limit the potential impact of an exposed credential.
The company also says it now checks all pull requests for credentials, has made zero data retention (ZDR) the default for new accounts, and has updated its processes for using AI coding agents. Customers handling sensitive prompt data are advised to enable ZDR in their account settings. Together AI directs customers with questions to security@together.ai.
Why it matters
Customers sending sensitive prompts to Together AI do not yet know from the company’s announcement whether affected samples contained personal or other customer data. Together AI advises customers handling sensitive prompt data to enable ZDR in account settings.
What it means for you
A customer handling sensitive prompt data should check account settings and enable ZDR; Together AI says it is on by default for new accounts, but does not say that existing accounts have been changed.
Our take: Together AI has described a specific route into the stored samples and listed concrete security changes. But customers still have to wait for the ongoing review to learn whether their prompt data included sensitive information.