Together AI says prompts sent to its serverless inference service leaked
Together.ai: Prompt Leakage for Requests
Together AI disclosed that an unauthorized party downloaded samples of prompts submitted to its serverless inference service in June and July 2026. The samples involved MiniMax M3, GLM-5.2 and Kimi K2.6. The leak began when an AI coding agent committed credentials for a private Hugging Face repository to the public GitHub repo xorl, and the unauthorized party also searched for and tested other service credentials in the samples.
Why it matters: An AI coding agent committed credentials to a public GitHub repo, and the window before the commit was reverted was long enough for samples to be downloaded and their credentials tested.