Together AI says coding agent leaked key that exposed customer prompts
An attacker downloaded stored prompt samples from June and July requests, then tested credentials found inside them, Together AI said.
By VibeLeaderboard
Nov 22, 2026, 12:16 AM UTC
Together AI disclosed that an unauthorized party downloaded a subset of prompts that customers sent to its serverless inference service in June and July 2026. The company traced the access to a credential that an AI coding agent had committed to a public code repository.
According to the notice, the prompts were samples Together AI had collected for routine software testing and stored in a private Hugging Face repository. They came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, mostly on June 24, July 4, July 8 and July 14.
The leak began in the company's own development work. While building its public GitHub repository, xorl, an AI coding agent committed a credential that opened the private Hugging Face repository. Together AI said its automated scanning caught the credential and the commit was reverted, but the unauthorized party obtained it during the exposure window.
The party then searched the downloaded prompts for credentials to other services and tested whether they worked, the company said. Five of the credentials it tested were canary credentials tied to a Dragos security experiment. Together AI said it worked with an external forensic firm and law enforcement, found no evidence of prompt leakage beyond this incident, and understands the data has since been deleted.
The review is not finished. Together AI said it is still examining the affected prompts to determine whether sensitive personal information or other customer data was exposed, and will contact affected customers directly.
The company listed its changes: scoped or short-lived credentials, with keyless access where supported; credential checks on every pull request; zero data retention (ZDR), meaning prompts are not stored, as the default for new accounts; and updated rules for using AI coding agents. It advised customers handling sensitive prompt data to enable ZDR in their account settings.
Why it matters: two ordinary practices combined here. Prompts kept for testing became a store of customer data, and an agent with commit access published the key to that store. Because the party searched the samples for credentials and tested them, any secret a customer had pasted into a sampled prompt was exposed along with it, not just the prompt text.
Checking sign-in…
Loading comments…