An AI coding agent leaked a credential that exposed customer prompts at Together AI
Together.ai: Prompt Leakage for Requests

Together AI is notifying customers after an unauthorized party downloaded stored prompt samples from its serverless inference service, collected in June and July 2026 from the MiniMax M3, GLM-5.2, and Kimi K2.6 models. The credential was committed to a public GitHub repository by an AI coding agent during development of the xorl project, then used to access a private Hugging Face repository. The intruder searched the samples for credentials to other services, testing five canary credentials from a Dragos security experiment. Together AI has since adopted scoped credentials, mandatory pull request scanning, and zero data retention by default.
During development of our public GitHub repository, xorl, an AI coding agent committed a credential that provided access to the private Hugging Face repository.
- tczMUFlmoNk
> During development of our public GitHub repository, xorl, an AI coding agent committed a credential that provided access to the private Hugging Face repository
My emphasis on the accountability laundering. No, "xorl" did not do this—you did. And, if "xorl" is https://github.com/togethercomputer/xorl and `togethercomputer` is together.ai, then you wrote "xorl" and are doubly responsible.
If five years ago I had said, "During development of our public GitHub repository, my computer decided to commit a credential", this would have been immediately recognized for the embarrassing abdication of duty that it is. So, too, should this be now.
- sans_souse
Am I inferring correctly here, in that Dragos basically had some sort of experiment set up and running for exactly this scenario, that effectively allowed them to trace the breach to the individual?
- ChrisArchitect
Related:
Data Processing Incidents on Inference Providers