Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Together AI Discloses Prompt Leak After AI Coding Agent Exposed Repository Credential

Together AI said on November 21 that an unauthorized party downloaded stored samples of customer prompts from its serverless inference service. Access traced back to an AI coding agent that committed a credential for a private Hugging Face repository while the company was building its public xorl repo, and that credential was used before the leak was contained.

Published

Together AI disclosed on November 21, 2026 that an unauthorized party downloaded stored samples from a subset of customer prompts submitted to its serverless inference service. The prompts had been authorized for routine software testing and were kept in a private Hugging Face repository, which is where the intruder found them.

The access path runs through AI-assisted development. While building its public xorl GitHub repository, an AI coding agent committed a credential that granted access to that private Hugging Face repository. Together AI's automated scanning detected the exposure and reverted the commit. Detecting and reverting that commit did not stop the credential from being used during the exposure window; an unauthorized party obtained it and used it to download the stored samples.

The downloaded samples map to specific traffic. The requests went mainly to MiniMax M3, GLM-5.2 and Kimi K2.6, on June 24, July 4, July 8 and July 14, 2026. That spread suggests a retained test archive rather than a single day's debugging session.

The intruder did more than read the prompts. After the download, the party searched the samples for credentials to other services and tested whether those credentials were valid. Five of the tested credentials were canary credentials tied to a Dragos security experiment, the kind of decoy designed to register when someone tries to use it.

Together AI said it worked with an external forensic investigation firm and law enforcement. Its review has not identified evidence of additional prompt leakage beyond this incident, the company said, adding that it has no reason to believe the data was leaked further and that it understands the data has now been deleted.

Customer impact still depends on what the samples contained. The review is ongoing to determine whether sensitive personal information or other customer data was exposed, and Together AI will contact affected customers directly with findings relevant to them.

The remediation list is split between credentials and process. Together AI is moving to scoped or short-lived credentials, keyless access where supported, credential checks on all pull requests, and updated development rules for AI coding agents. Zero data retention is now the default for new accounts, and the company recommends that customers handling sensitive prompts turn it on.

The incident matters because two risk chains meet here: credential hygiene in AI-assisted development, and the data surface created when customer prompts are retained for testing. Together AI has documented controls for both, but the content review now under way is what will determine whether this is a contained credential lapse or a customer data incident.

Why it matters

For enterprises using hosted inference, the takeaway is that prompt retention policy and developer-side credential management are one security problem, not two.

Together AISecurityCoding Agent
Back to realtime news

Nearby Updates

All