Together AI disclosed a security breach involving prompt samples from its serverless inference service in June and July 2026, caused by credential exposure.
Together AI has disclosed a security incident involving the unauthorized access and download of prompt samples from its serverless inference service during June and July 2026. The breach occurred after a credential for a private repository was inadvertently exposed in a public GitHub repository. The company is currently investigating the scope of the incident and notifying affected customers as necessary.
According to Together AI, the issue stemmed from routine software testing that involved storing authorized prompt samples in a private Hugging Face repository. These samples originated from requests to the MiniMax M3, GLM-5.2, and Kimi K2.6 models, primarily on June 24, July 4, July 8, and July 14, 2026. During the development of the public GitHub repository "xorl," an AI coding agent inadvertently committed a credential granting access to the private repository.
The exposed credential was detected by automated scanning tools, and the commit was promptly reverted. However, during the exposure window, an unauthorized party accessed the repository and downloaded the stored prompt samples. The party reportedly searched the samples for additional credentials and tested five canary credentials associated with a Dragos security experiment. Together AI stated that it has no evidence of further data dissemination and believes the unauthorized party has since deleted the data.
Together AI has engaged an external forensic investigation firm and law enforcement to assess the breach. While no additional prompt leakage has been identified beyond this incident, the company is still reviewing the affected data to determine whether sensitive personal information or other customer data was exposed.
Source link







