Sulat.com / Writing on AISunday, November 22, 2026
Artificial Intelligence/22 November 2026

Together AI says a coding agent exposed a credential, and someone downloaded stored prompts

Together AI says an unauthorised party used a credential an AI coding agent committed to its public xorl repo to download stored prompts sent to MiniMax M3, GLM-5.2 and Kimi K2.6. Five credentials it tested were Dragos canaries.

Glossy 3D Together AI logo, three coloured discs, in a data centre aisle beside an open vault spilling glowing pages, with a brass key hanging above them

At a glance

  • Together AI says an unauthorised party downloaded stored prompt samples from its serverless inference service after an AI coding agent committed a credential to its public xorl repository.
  • The samples came from requests to MiniMax M3, GLM-5.2 and Kimi K2.6, primarily on 24 June and 4, 8 and 14 July 2026.
  • The notice says five of the credentials the party tested were canary credentials associated with a Dragos security experiment.
  • Zero data retention is now the default for new Together accounts, and Together tells customers with sensitive prompts to switch it on.
  • The notice doesn't give the number of customers or prompts involved, or when the credential was committed.

Together AI says an unauthorised party downloaded stored prompt samples from its serverless inference service after an AI coding agent committed a credential to a public GitHub repository. CEO Vipul Ved Prakash posted the company’s security notice on 21 November. The samples came from requests to MiniMax M3, GLM-5.2 and Kimi K2.6, mostly on four days in June and July 2026.

The repository was xorl, the public home of Together’s distributed training framework for large language models. The credential opened a private Hugging Face repository where Together kept prompt samples it had collected for routine software testing. Together’s scanner flagged the credential and the commit was reverted. The notice says the unauthorised party got the credential while it was exposed and used it to download the samples.

Vipul Ved Prakash, Together AI’s co-founder and CEO, announcing the notice on X, 21 November.

What the notice says

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

QuestionWhat Together’s notice says
What was takenStored prompt samples from its serverless inference service
Which modelsMiniMax M3, GLM-5.2 and Kimi K2.6
Which daysPrimarily 24 June, 4 July, 8 July and 14 July 2026
How it got outAn AI coding agent committed a credential to the public xorl repository
Where the samples satA private Hugging Face repository
Where the data is nowTogether understands it has been deleted
Source: Together AI’s security notice, 21 November.
Five-step chain: coding agent commits a credential, it opens a private repo, Together reverts it, an unauthorised party downloads samples, five tested credentials were canaries
The credential was caught and reverted, but not before someone used it.

Five of the tested credentials were canaries

After the download, the notice says, the unauthorised party 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. A canary credential is a fake that exists in one place only, so any attempt to use it shows that place was read.

Dragos published the results of that kind of experiment on 19 November. HD Moore, its chief research officer, wrote that Dragos sent nearly two million canary-seeded prompts to AI services from July through November and found evidence that data from at least three providers had been exposed. Dragos’s post doesn’t name Together or any other provider, and Together’s notice doesn’t say whether it’s one of the three. Three of Together’s four main sample dates fall in July, the month Dragos says it started seeding prompts.

Three stat cards: 3 models sampled, 4 main sample dates, 5 canary credentials tested; a note says customer and prompt counts aren't given
Three models, four days and five canaries are the only counts in the notice.

What Together has changed

The notice lists four changes. Together says it now uses scoped or short-lived credentials, and keyless access where supported, to limit the damage when one is exposed. It checks every pull request for credentials. It has updated its development processes for AI coding agents. And zero data retention, ZDR, is now the default setting for new accounts.

Together’s privacy docs say it stores the prompts you send and the responses models return by default, and that the data isn’t shared with third parties. An organisation admin switches storage off under Organization Settings, in the Privacy section, which turns ZDR on. The ZDR page says it applies only to requests sent after it’s enabled, and that turning off prompt storage also turns off passthrough models, the ones that forward requests to a third-party provider.

Two cards: new Together accounts have ZDR on by default; existing accounts switch it on under Organization Settings, Privacy, by turning off prompt storage
New accounts start with ZDR on; existing ones have to switch it on.

What the notice leaves out

The notice doesn’t say how many customers or prompts were in the samples, when the credential was committed or how long it stayed exposed. It doesn’t say who the unauthorised party is, which customers Together has contacted, or whether prompts sent to other models were stored the same way. Together says it understands the data has now been deleted, and doesn’t say who confirmed that.

If you sent prompts to those models

Together’s notice doesn’t tell customers to rotate keys. The unauthorised party searched the samples for credentials and tested them, so any secret that sat in a prompt from those dates is worth replacing.

  1. Check your Together usage for MiniMax M3, GLM-5.2 and Kimi K2.6 requests around 24 June and 4, 8 and 14 July.
  2. Look inside those prompts for API keys, tokens, database URLs and passwords.
  3. Revoke and replace anything secret you find. A new key issued next to the old one leaves the old one live.
  4. Turn off prompt storage in Organization Settings, Privacy, to switch ZDR on for later requests.
  5. Ask Together at [email protected] if you need to know whether your organisation was affected.

This is separate from OpenRouter’s email of 21 November, which told customers that a third party may have read prompts sent to MiniMax models. In that case a researcher’s test links were opened days after the requests, and the email names AtlasCloud, Novita and StreamLake as the forwarding providers. It doesn’t name Together.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.