Together AI Discloses Prompt Leak After Coding Agent Exposed Credential

What happened
Together AI announced on November 21, 2026, that an unauthorized individual downloaded stored samples of customer prompts from its serverless inference service. The incident occurred after an AI coding agent exposed a private credential during the development of the public xorl GitHub repository. The exposed credential allowed access to a private Hugging Face repository where prompt samples were stored for software testing. Together AI began a forensic investigation and is reviewing the extent of any data exposure.
Key details
- The exposed credential granted access to prompt samples involving MiniMax M3, GLM-5.2, and Kimi K2.6 models.
- Prompt samples were primarily from June 24, July 4, July 8, and July 14, 2026.
- Five credentials tested by the unauthorized party were canary credentials from a Dragos security experiment.
- Together AI has implemented scoped or short-lived credentials, credential checks on pull requests, and made zero data retention the default for new accounts.
- The company is working with law enforcement and an external forensic firm to investigate the incident.
Why it matters
Customers whose prompts were stored for testing may have had their data accessed, depending on the content of those prompts. Organizations using Together AI's services now face new security measures, including default zero data retention and stricter credential controls, which may affect how they manage and store prompt data. Affected customers will be contacted directly if their information was exposed, and they may need to review their own security practices in response to the incident.
