ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI compliance
Together AI Prompt Leak: Self-Check Guide for June–July Users of Three Models

Together AI Prompt Leak: Self-Check Guide for June–July Users of Three Models

AI compliance • Admin • • 6 views

On November 21, 2026, Together AI announced on its official blog that prompt samples submitted to its serverless inference service in June and July 2026 had been downloaded by an unauthorized party. The users mainly affected are those of three models — MiniMax M3, GLM-5.2 and Kimi K2.6 — with requests concentrated on June 24, July 4, July 8 and July 14. The company says it has brought in an external forensic firm and law enforcement.

How the leak happened

In routine testing, Together AI collected authorized prompt samples and stored them in a private Hugging Face repository. The failure came during development: while working on xorl, a public GitHub repository, an AI coding agent committed a credential that granted access to that repository. Automated scanning spotted it and the commit was reverted, but within the exposure window an unauthorized party had already used the credential to download the samples, then searched them for credentials to other services and tested them. Five of those credentials were canary credentials planted for a Dragos security experiment. Together AI says the data was not leaked further and the other party has deleted it.

Who is in scope

The leak has clear boundaries: only the serverless inference service, only the three models above, and requests concentrated on four dates. But whether you are affected is not something users can fully confirm on their own yet — Together AI is still reviewing the affected prompts one by one to confirm whether they contained sensitive personal information or other customer data; the review is ongoing, and the company will contact the relevant customers directly. Users can start by checking their own records: did you use any of these three models in June or July, and did your prompts contain API keys, personal information or business material? How to view, delete and manage saved chat content may help with that self-check. The notice does not disclose how many users were affected or where they are located, and breach-notification rules differ by jurisdiction, so one region's rules cannot be applied to every user.

Three things to do now

First, turn on zero data retention (ZDR) in your account settings: Together AI has made it the default for new accounts, and users handling sensitive prompt data should enable it. Second, go back over the prompts you sent to the three models in June and July: rotate any API key, password or access token that appeared in them, and assess the impact if personal information or business secrets appeared. Third, watch for a direct notice from Together AI, and raise questions through its official security contact channel. The company has since changed four practices, including short-lived credentials and ZDR by default for new accounts. This is not the only recent case of its kind — Meta's Muse has also been drawn into a dispute over exposed internal chats and private-message access.

Two risks remain open: credentials found in the prompts were actually tested by the other party, and whether sensitive personal information was included is still under review. The boundaries matter too: the leak covers only the serverless service, the three models and the four dates — not every Together AI user. The lesson is not about the models themselves: putting secrets into prompts and committing credentials into a code repository are two ordinary habits, and together they turned a stored test sample into a real data leak.

Recommended Tools

More