An AI coding agent committed a credential to a public GitHub repository, and an intruder used it to download customer prompts stored by Together AI's serverless inference service. Together AI disclosed the incident on November 21, 2026, and reporting on it carried into Sunday. This Together AI prompt leak is a follow-on to the prompt-privacy story USA 24 has already covered, but it adds a new twist: the careless commit came from software, not a developer.
How the credential reached a public repository
The leak started in Together AI's public GitHub repository, called xorl. An AI coding agent working on that repository committed a credential that granted access to a private Hugging Face repository, and that private repository held prompt samples. StartupHub reported the disclosure, and a DEV Community write-up walked through the sequence. The details here come from those secondary reports, not from Together AI's own disclosure post.
Together AI says its automated scanning caught the exposed credential and reverted the commit. By that point, an unauthorized party had already downloaded data during the exposure window. The reports do not say how long the credential sat in public view before the scan flagged it. Scanners and scrapers both watch public repositories, so the time between a push and a takedown is the real risk window.
What the Together AI prompt leak exposed
According to explainx.ai's summary, the exposed prompts were requests sent to three models during June and July 2026:
- MiniMax M3
- GLM-5.2
- Kimi K2.6
None of the sources reviewed disclose how many customers were affected. Together AI says it is working with an external forensic firm and law enforcement while it reviews the exposed prompts for sensitive personal information or other customer data, according to Fox News' live coverage. That review will decide who has to be notified.
Prompts are often where people paste the most sensitive material. A developer debugging a failing service may paste an error log that contains a token, and a support team may paste a customer record into a model to get a summary. Anyone who sent that kind of material to these models in June or July should treat the question as open until the company's review is finished.
What the intruder did with the stolen prompts
The intruder did not stop at reading the files. According to the reports, it searched the downloaded prompts for credentials to other services and tested whether those credentials worked.
Five of the tested credentials were canary traps, planted as part of a Dragos security experiment. Canary credentials are decoys that reveal their use, so the intruder's tests left a record of what it tried. That record could help investigators trace the activity, though the reports do not say whether the canary tokens were triggered anywhere else.
No source reviewed names a suspect or a group behind the activity. Together AI says it has no reason to believe the data spread further, and it understands the downloaded copy has been deleted. Those are the company's statements, and outside observers cannot check either claim yet.
How the Dragos research and OpenRouter cuts fit in
The leak arrived days after Dragos published research on November 19 showing that prompts sent to AI services could be stolen, logged and passed around. A DEV Community post summarized that finding for developers.
The same weekend brought a second signal. On November 22, OpenRouter said it had removed three providers and stripped 125 endpoints of their zero data retention (ZDR) status. OpenRouter's public list links notices from Novita, Phala and Together AI, but it does not say which three providers were removed.
Together AI's disclosure helps explain part of that gap, because it names a root cause for the kind of exposure that OpenRouter's removals had left unexplained. The sources do not confirm which three providers OpenRouter cut, and not every affected provider has published a notice. Treat the two events as related, not as one confirmed story. Hosted AI services are becoming a recurring point of exposure, and USA 24 has covered the Fireworks AI breach, which reached Uber drivers and Cursor and Notion users.
What to do if your prompts went to Together AI or another hosted model
Work through this checklist:
- Watch for a direct notice. If you used Together AI's serverless inference in June or July 2026, the company says it will contact affected customers directly.
- Review the prompts you sent during that period. Look for API keys, tokens, passwords, customer records and personal data.
- Rotate every credential that appeared in those prompts. Do this for any hosted model provider you used in the window, not only Together AI.
- Enable zero data retention where the provider offers it. Together AI has made ZDR the default for new accounts.
- Scan repositories and pull requests for secrets before merging AI-agent commits. Together AI says it now runs credential checks on every pull request.
- Give coding agents scoped, short-lived credentials instead of long-lived tokens.
If your company holds customer data, bring in legal counsel before deciding whether notification is required. The answer depends on what the prompts contained and on the state laws that apply to that data, so a developer's guess is no substitute for that review.
Why AI coding agents change the secret-leak risk
Leaking secrets through public code is an old problem. Developers have pushed API keys and passwords into public repositories for years, and scanners exist to catch them. What changes here is the committer. The agent working on the xorl repository did not need to be malicious for the exposure to happen. A blog post on the incident makes the same point, and the mechanism is simple: an automated tool with access to a secret can write that secret into a commit while doing its job.
The threat data points the same way. Beazley Security's third-quarter 2026 report, released November 19, found that reported ransomware incidents rose about 38% from the second quarter. According to its press release, the share of cases that started with an exploited vulnerability doubled to 28%. The full quarterly report also describes an intrusion carried out largely by coordinated AI agents. Those figures come from one vendor's clients, not a market-wide count, so they show direction more than scale.
What to watch next
Three things bear watching. First, Together AI's full disclosure should be checked against the secondary reports used here, and the company's count of notified customers will show how widely the prompts were copied. Second, OpenRouter's three removed providers remain unnamed, which leaves the ZDR picture incomplete. Third, the Dragos canary tokens could turn up in other incidents, which would show whether this intruder's testing reached beyond one provider.
The larger question is whether zero retention, pull-request scanning and scoped agent credentials become the default across hosted AI services, or stay optional settings that only careful teams switch on. Until that changes, the safest assumption is that a prompt can outlive the session that created it.




