Together AI Discloses Prompt Leakage Incident Caused by AI Coding Agent Credential Exposure
Together AI CEO disclosed a security incident where an AI coding agent committed a credential, exposing customer prompts from June and July 2026.
AI infrastructure company Together AI has disclosed a security incident involving the leakage of stored prompt samples submitted to its serverless inference service. The breach occurred in June and July 2026 but was publicly announced on November 21, 2026, by CEO Vipul Ved Prakash.
The root cause was an error by an AI coding agent during development. While working on the public GitHub repository 'xorl', the agent committed a credential that provided access to a private Hugging Face repository containing authorized prompt samples. These samples originated from requests to MiniMax M3, GLM-5.2, and Kimi K2.6 models. An unauthorized party obtained this exposed credential and downloaded the stored prompts.
During the exposure window, the attacker searched the leaked prompts for other service credentials and tested their validity. Five of these were canary credentials associated with a Dragos security experiment. Together AI stated it has no reason to believe the data was leaked further and understands it has been deleted.
In response, the company has strengthened controls around development workflows and credentials. Key changes include using scoped or short-lived credentials, checking all pull requests for secrets, making zero data retention (ZDR) the default for new accounts, and updating processes governing the use of AI coding agents. Customers handling sensitive data are advised to enable ZDR.