If you send work emails, customer details, or internal documents to an AI service, you probably assume those messages stay private. This announcement is a reminder that where your data is stored matters as much as how the AI answers you.
What happened
According to Together AI's announcement, the company collected some prompt samples during routine software testing and saved them in a private storage location on Hugging Face, a platform where developers share code and AI models. A prompt is simply the text you type into an AI tool, and a sample is a copy of some of those messages.
The samples came from requests sent to three open models: MiniMax M3, GLM-5.2, and Kimi K2.6. Most of the affected requests date from June 24, July 4, July 8, and July 14, 2026.
The company explains that an AI coding tool used during development of a public project committed a credential, which is a digital key that proves who you are and grants access. That key opened the private storage location. Together AI's automated scanning spotted the exposed key, and the company reverted the change, but an outside party had already found the key and used it to download the samples.
The announcement also says that party searched the samples for other passwords and keys, and tested whether some of them still worked. Five of the tested credentials were deliberate test keys tied to a security research experiment. Together AI says it has no reason to think the data spread further and understands the copies have been deleted.
What the company says it has done
Together AI says it worked with an outside forensic investigation firm and with law enforcement. It reports that its review has not found evidence of other prompt leaks beyond this incident.
The company also lists changes it has made:
- Using limited or short-lived credentials, and keyless access where the system supports it, so a leaked key does less damage.
- Checking all code changes for credentials before they go live.
- Making zero data retention (ZDR) the default setting for new accounts. ZDR means the provider does not keep your prompts after your request is handled.
- Updating its rules for how AI coding tools are used in development.
Together AI also says its review of the affected prompts is still ongoing, to find out whether personal information or other customer data was included. It will contact affected customers directly if it finds something relevant to them. The company's CEO, Vipul Ved Prakash, apologized in the announcement.
Wrap-up
This incident shows how one leaked access key can expose data that was never meant to leave a company's systems. Together AI says it is still reviewing what was affected, and it will reach out to customers when it has specifics. Your next step today is simple: look at how your organization's AI accounts handle stored prompts, and decide whether any of them need ZDR or a more careful rule about what goes in.
