Coding & agents New
Together AI says a coding agent leaked a credential that let an outsider download stored customer prompts
Published:
Together AI, which runs open models for developers as a cloud inference service, disclosed on Saturday that an unauthorized party downloaded stored samples of customer prompts sent to its serverless inference service in June and July 2026. The samples came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models and were kept in a private Hugging Face repository for software testing, the company says.
Together AI gives the cause: during development of its public GitHub repository xorl, an AI coding agent committed a credential that gave access to that private repository. Automated scanning caught it and the commit was reverted, but during the exposure window someone obtained the credential, downloaded the samples, searched them for credentials to other services and tested whether those worked.
Five of the credentials the party tested were canary credentials from a Dragos security experiment, Together AI says. Dragos published that study on Thursday 19 November: it says it sent about two million prompts carrying unique, trackable secrets to AI services between July and November and saw evidence that data from at least three providers had been exposed, without naming them.
Together AI says it worked with an outside forensic firm and law enforcement, believes the data has now been deleted, and is still reviewing which customers to notify. Its changes include scoped or short-lived credentials, credential checks on every pull request, zero data retention by default for new accounts, and new rules for developing with AI coding agents.
Caveat. The account of the cause and scope is Together AI's own. Its post gives a date but no time, and does not say how many customers or prompts were affected or how long the credential was exposed. Dragos did not name the three providers in its study; the link between the two rests on Together AI's statement.
Sources