Skip to content
Breach Index · running record

Together AI security incident in 2026: prompt samples downloaded after a leaked credential

Together AI runs a serverless inference service for open models. This page tracks the security event it disclosed on 21 November 2026, in which stored samples of customer prompts were downloaded by an unauthorized party.

Short answer

Was Together AI hacked?

Yes, by Together AI's own account: an unauthorized party downloaded stored prompt samples. Together AI says that while it was developing a public GitHub repository, "an AI coding agent committed a credential that provided access to the private Hugging Face repository" where it kept prompt samples for routine software testing. Its scanning caught the credential and the commit was reverted, but the party had already obtained it.

The samples came from requests to MiniMax M3, GLM-5.2 and Kimi K2.6 on its serverless inference service, mainly on 24 June and 4, 8 and 14 July 2026. Together AI says the party searched them for credentials to other services and tested them; five were canary credentials from a Dragos security experiment. It is working with an external forensic firm and law enforcement.

Dragos's 19 November research describes three breached AI providers but names none of them; no source says whether Together AI is one of them.

Timeline at a glance

  1. 24 Jun – 14 Jul 2026Main dates of the serverless requests (MiniMax M3, GLM-5.2, Kimi K2.6) whose stored samples were later downloaded. (Together AI)Confirmed
  2. 2026 (date not given)An AI coding agent commits a credential to a public GitHub repository; scanning detects it and the commit is reverted. (Together AI)Confirmed
  3. 2026 (date not given)An unauthorized party downloads the prompt samples, searches them for credentials and tests them; five are Dragos canaries. (Together AI)Confirmed
  4. 19 Nov 2026Dragos publishes its canary research on breached AI providers (providers unnamed). (Dragos)Reported
  5. 21 Nov 2026Together AI publishes its security notice; ZDR becomes the default for new accounts. (Together AI)Confirmed
  6. 22 Nov 2026OpenRouter links Together AI's notice in its announcement on inference-provider incidents. (OpenRouter)Reported

Last updated 22 November 2026. Every entry is dated and linked to its source; claims by attackers are labelled as claims, never stated as fact.

Together AI incident timeline, in detail

Confirmed = acknowledged by Together AI, a regulator filing or a court. Reported = credible reporting Together AI has not fully confirmed. Unverified claim = an attacker or leak-site claim.

Jun–Nov 2026ConfirmedCredential leak / unauthorized data download

Credential committed by an AI coding agent exposes stored prompt samples

What happened. Together AI says it "collected prompt samples authorized for that use and stored them in a private Hugging Face repository" for routine software testing. While developing its public GitHub repository xorl, an AI coding agent committed a credential that gave access to that repository. Automated scanning detected it and the commit was reverted; Together AI does not give the dates.

Impact. An unauthorized party used the credential to download the samples (requests to MiniMax M3, GLM-5.2 and Kimi K2.6, mainly 24 June and 4, 8 and 14 July 2026), searched them for credentials to other services and tested whether they worked. Together AI says it has found no prompt leakage beyond this incident, is still reviewing whether personal or other customer data was exposed, and will contact customers directly as needed.

Sources: Together AI notice · PlayCISO write-up

Where the reports disagree

  • Dates: Together AI gives the dates of the sampled requests but not when the credential was committed, detected or used.
  • Link to the Dragos research: Together AI says five of the tested credentials were Dragos canaries. Dragos names none of the three providers it found breached, and OpenRouter does not say Together AI is one of the three providers it off-boarded.
  • Scope: Together AI calls it "a subset of prompts" and has not said how many customers or requests are affected.

The pattern across the record

AI coding agents leak secrets too

The credential was committed by an AI coding agent working on a public repository. Agents need the same secret scanning, scoped credentials and review as people.

Test data is production data

Prompt samples kept for testing are customer data. A copy in a developer-facing store is as sensitive as the live service.

Stolen prompts are mined for keys

The party searched the samples for credentials and tested them, which is how the Dragos canaries surfaced.

What it means for Together AI customers — and what to do

If you sent prompts to MiniMax M3, GLM-5.2 or Kimi K2.6 on Together AI's serverless service in June or July 2026, treat any secret in them as exposed.

  1. 1.Rotate secrets sent in prompts

    Any credential in requests to those models in June and July 2026, then review the accounts for misuse.

    Secret rotation order planner
  2. 2.Turn on zero data retention

    Together AI advises customers handling sensitive prompt data to enable ZDR in account settings; it is now the default only for new accounts.

  3. 3.Scan what your own agents commit

    Check every pull request, including agent-written ones, for secrets, and give agents short-lived, scoped credentials.

    Secret scanner
Take this with you
Incident timeline · What Together AI customers should do · Sources

CISO takeaways

Sources

Company statements and regulator filings first, then reputable reporting. Reviewed 22 November 2026.

  1. 1.Security event: prompt leakage for requests in June and July of 2026 — Together AI, 21 Nov 2026 (updated 22 Nov)
  2. 2.Data Processing Incidents on Inference Providers — OpenRouter, 22 Nov 2026
  3. 3.Three breaches, endless leaks: Your AI supply chain needs canaries — Dragos, 19 Nov 2026

Frequently asked questions

Was Together AI hacked?

Together AI says an unauthorized party downloaded stored prompt samples after an AI coding agent committed a credential to a public GitHub repository, giving access to the private Hugging Face repository that held them.

Which Together AI prompts were exposed?

Samples of requests to MiniMax M3, GLM-5.2 and Kimi K2.6 on its serverless inference service, mainly from 24 June and 4, 8 and 14 July 2026, per Together AI.

What did Together AI change?

It made zero data retention the default for new accounts, moved to scoped or short-lived credentials, checks every pull request for credentials and updated its processes for AI coding agents.

Is Together AI one of the three providers Dragos found breached?

No source says so. Together AI says five tested credentials were Dragos canaries, but Dragos does not name its three breached providers.

Same sourced, verification-tagged timeline, company by company — compare how peers have fared.

Or see the full Breach Index and how to read it.

Independent analysis for security-leadership education; not affiliated with or endorsed by Together AI. Facts are attributed to the dated sources above, and anything disputed or unconfirmed is labelled as such. See the full Breach Index.