Skip to content

AI NewsPublished 6 min read

Prompt Leak, Together AI Reports; ZDR to Weigh

nodes passing glowing task tokens along branching paths
Listen to this article · 9:12 · AI-generated narration
0:00 / 9:12
Chapters

Together AI Reports Prompt Samples Were Taken

Together AI said on November 21, 2026, in a post on its Together company blog, that an unauthorized party downloaded stored samples of prompts submitted to its serverless inference service in June and July 2026. The samples came primarily from requests sent on June 24, July 4, July 8, and July 14, 2026, according to Together AI. The download followed the exposure of a credential for a private repository in a public code repository, the company said.

Customers had authorized the prompts for routine software testing, and Together AI stored the samples in a private Hugging Face repository, Unite.AI reported on November 21, 2026. The company apologized in its notice and said it is notifying customers about the event. Together AI described the affected material as a subset of the prompts sent to the service.

The short version

Together AI said an unauthorized party downloaded stored samples of customer prompts sent to its serverless inference service in June and July 2026, after an AI coding agent committed a private-repository credential to a public GitHub project. The company is still reviewing whether the samples held sensitive personal information or other customer data, and said it will notify affected customers directly as needed.

  • Five Dragos canary credentials were among those the intruder tested, Together AI said.
  • An external forensic firm and law enforcement worked on the case, and Together AI said it understands the data has been deleted.
  • New accounts now default to zero data retention, and Explainx reported existing accounts must switch it on themselves.
  • OpenRouter listed the Together AI notice among provider security announcements it received after a Dragos canary study.

A coding agent committed the access key

During development of xorl, the company's public GitHub repository, an AI coding agent committed a credential that provided access to the private Hugging Face repository, Together AI said. Automated scanning detected the exposed credential, and the company reverted the commit, according to the post. In the exposure window, an unauthorized party obtained the credential and used it to download the stored samples, the company said.

Reverting the commit did not stop the credential from being used during the exposure window, Unite.AI reported. Explainx reported on November 22, 2026, that xorl is tied to Together AI's open inference stack and that a revert does not revoke a credential that was already copied.

The intruder searched prompts for secrets

After the download, the unauthorized party searched the samples for credentials to other services and tested whether those credentials were valid, Together AI said. Five of the credentials tested were canary credentials tied to a Dragos security experiment, according to Together AI's November 21 notice.

DEV Community explained on November 22, 2026, that a canary token is a fake secret placed in one location only, so any use of it shows that location was exposed. Dragos built its canaries so every hit could be tied back to the exact prompt, model, router, and privacy setting that carried it, DEV Community reported. The intruder's activity extended to possible reuse of secrets contained in prompts, a risk beyond disclosure of the prompt text itself, Unite.AI reported.

Together AI's data review is still open

Together AI said it worked with an external forensic investigation firm and law enforcement. Based on its review, the company said it has not identified evidence of additional prompt leakage beyond the June and July event. It has no reason to believe the unauthorized party leaked the data further and understands the data has now been deleted, the company said.

Its review of the prompt contents continues, to determine whether sensitive personal information or other customer data was exposed, and Together AI said it will contact affected customers directly as needed. The company does not say how many distinct customers appear in the archive, Explainx reported. Customer impact still depends on the content of the samples, Unite.AI reported.

Together AI tightens credential and data rules

Following the breach, Together AI said it strengthened controls around development workflows, credentials, and customer data, listing the changes below.

Together AI, which runs the serverless inference service, wrote in its notice: "Customers handling sensitive prompt data should enable ZDR in their account settings." Zero data retention tells the platform not to retain prompt and completion text for training or logging beyond what operations require, Explainx reported. ZDR governs production inference paths and does not automatically cover engineering datasets copied into private Hugging Face repositories for QA, according to Explainx. Existing accounts must still change the setting themselves, Explainx said.

AreaChange Together AI said it made
CredentialsUses scoped or short-lived credentials, and keyless access where supported
Code reviewChecks all pull requests for credentials
Data retentionMakes zero data retention (ZDR) the default for new accounts
AI coding agentsUpdated development processes governing their use

Dragos and OpenRouter moved in the same week

Dragos published a canary study on November 19, 2026, in which HD Moore, its chief research officer, and his team sent nearly two million prompts to AI services between July and November 2026, DEV Community reported. HD Moore described AWS credentials, planted only in requests a provider rejected before a model ran, that were later used, Penligent reported on November 21, 2026.

After reviewing the findings, OpenRouter said on November 22, 2026, that it off-boarded three providers and removed 125 endpoints from Zero Data Retention eligibility. The findings showed no breach or unauthorized access of OpenRouter itself, OpenRouter said. Notices from Novita, Phala, and Together appear in OpenRouter's live list of provider security announcements. Explainx placed the Together AI disclosure in the same week as the Dragos research and OpenRouter's MiniMax customer emails.

Tron's take

My take: the useful lesson in the Together AI notice is about copies. Customers sent prompts for inference, the company kept a sample set for testing, and an AI coding agent exposed the key to that second store. Two common readings miss it. One says a business must adopt every new AI release the week it ships; the other says AI news is irrelevant to a small firm. I hold neither view. A small or mid-sized business does better applying AI tools it has already proven, while tracking what changed at its providers so it can adjust on its own schedule.

My advice for any firm using hosted models: ask each AI vendor in writing where copies of prompts live beyond the inference path, including test datasets, evaluation harnesses, and support exports. Turn on zero data retention for existing accounts rather than assuming the new-account default reached them. Keep passwords and API keys out of prompts, since the intruder here searched the samples for credentials. If staff or contractors use AI coding agents, I would require credential scanning on every pull request before code reaches a public repository, the control Together AI now says it runs. XL.net covered AI key abuse earlier this month in Wiz Ties Black October Breaches to AI Key Abuse.

Together AI has not finished reviewing what the samples contained, so the scope may still change. Whether prompt copies deserve more attention than provider choice is a judgment call; that is my reading of the news, not a reported result. XL.net sells security assessments and managed IT services, including reviews of credentials and code repositories like the one that failed in the Together AI case.

Questions I'd expect

Which Together AI customers were affected?

Together AI said the samples came from a subset of prompts sent to its serverless inference service in June and July 2026. Its review of the contents is ongoing, and the company said it will contact affected customers directly as needed.

How did the credential reach the intruder?

An AI coding agent working on xorl, Together AI's public GitHub repository, committed a credential for the private Hugging Face repository holding the samples. The company reverted the commit after automated scanning flagged it, but an unauthorized party obtained the credential during the exposure window, Together AI said.

Does zero data retention cover test copies of prompts?

Not automatically, according to Explainx, which reported that ZDR governs production inference paths rather than engineering datasets copied for testing. Together AI now applies ZDR by default to new accounts, and Explainx reported that existing accounts must change the setting themselves.

All AI news