Together AI has disclosed a security incident in which an unauthorized party downloaded stored prompt samples from its serverless inference service. The notice was published on November 21, 2026 under the name of Vipul Ved Prakash, the company's CEO and co-founder. Together says the samples came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, made primarily on June 24, July 4, July 8 and July 14, 2026.

According to the notice, Together had collected the samples for routine software testing, with authorization, and kept them in a private Hugging Face repository. While the team was developing its public GitHub repository, called xorl, an AI coding agent committed a credential that gave access to that private repository. Together's automated scanning detected the exposed credential and the commit was reverted, but the company says an unauthorized party obtained the credential during the exposure window and used it to download the samples.

The notice adds that the party then searched the samples for credentials to other services and tested whether they were valid. Five of the credentials it tested were canary credentials associated with a Dragos security experiment. Sulat, in a report by JP Caparas dated November 22, notes that Dragos published results of a canary-prompt experiment on November 19, but also points out that neither Dragos nor Together says Together is one of the providers involved, so the link should not be assumed.

Together says it worked with an external forensic investigation firm and law enforcement and has found no evidence of additional prompt leakage beyond this incident. It understands that the data has been deleted. Its review of the affected prompt contents is still under way to determine whether sensitive personal information or other customer data was exposed, and it will contact affected customers directly as needed.

The company lists four changes: scoped or short-lived credentials and keyless access where supported, credential checks on all pull requests, zero data retention as the default setting for new accounts, and updated development processes for the use of AI coding agents. Customers who handle sensitive prompt data are advised to enable zero data retention in their account settings, and questions go to [email protected].

Independent coverage confirms the account but also notes the gaps. Unite.AI, in an article credited to its AI cybersecurity research agent Miles Okada and published on November 21, stresses that reverting the commit did not stop the credential from being used. Sulat and Oossa point out that the notice does not say how many customers or prompts were involved, when the credential was committed, how long it stayed exposed, or who the unauthorized party was.

For teams that send prompts to hosted inference providers, the incident is a reminder of two risks at once: retained prompt samples can become a target, and coding agents with repository access can leak secrets. Anyone who sent prompts to those three models on the named dates should check them for embedded keys and rotate any secret found, a step the press coverage recommends even though the notice itself does not.

References