Together AI has disclosed that customer prompts sent to its serverless inference service were downloaded by an unauthorized party — after one of its own AI coding agents committed a credential to a public code repository.

The company published the disclosure on November 21, 2026. The leaked credential unlocked a private repository holding prompt samples collected during routine software testing, and the exposure window was long enough for an attacker to take the lot before the commit was reverted.

How the leak happened

The samples came from requests to three models on Together AI's serverless inference service — MiniMax M3, GLM-5.2 and Kimi K2.6 — made primarily on June 24, July 4, July 8 and July 14, 2026. Together AI says the samples were collected with authorization for testing purposes and stored in a private Hugging Face repository.

During development of the company's public GitHub repository, xorl, an AI coding agent committed a credential that granted access to that private repository. Together AI says its automated scanning detected the exposed credential and the commit was reverted — but during the window it was live, an unauthorized party obtained the credential and used it to download the stored prompt samples.

What the intruder did next is what turned a credential slip into a wider incident. The party searched the downloaded samples for credentials to other services and tested whether the ones it found were valid. Five of the credentials it tested turned out to be canaries — decoy secrets planted as part of a security experiment by the industrial cybersecurity firm Dragos. Whether the intruder found any real credentials among the samples is still under review.

What Together AI is doing about it

Together AI says it has no reason to believe the unauthorized party leaked the data further, and understands the samples have now been deleted. It has hired an external forensic investigation firm, is working with law enforcement, and says it has found no evidence of additional prompt leakage beyond this incident. Its review of the affected prompt contents is still underway, to determine whether sensitive personal information or other customer data was exposed, and the company says it will contact affected customers directly.

Chief executive Vipul Ved Prakash apologized for the incident and reaffirmed the company's commitment to improving its security practices.

The company lists four changes it has made since the incident: moving to scoped or short-lived credentials, and keyless access where supported; checking all pull requests for credentials; making zero data retention the default setting for new accounts; and updating the development processes that govern the use of AI coding agents. Customers handling sensitive prompt data are advised to enable zero data retention in their account settings.

A leak the researchers saw coming

The disclosure lands days after Dragos published research showing how easily prompts can escape the services they are sent to. On November 19, Dragos chief research officer HD Moore described sending about two million prompts containing unique tracking links and decoy credentials to widely used AI services between July and November, then watching where the canaries surfaced. The study found three compromised providers and prompt-reading resellers — and Together AI's own disclosure confirms that five of the credentials its intruder tested were Dragos canaries.

Novita, another inference provider, published a privacy notice the same day, confirming some MiniMax requests had been forwarded to MiniMax's official API.

The episode is an awkward first for the industry: a coding agent — the kind of tool sold as a way to ship software faster — quietly committed the key that opened the door. For companies running inference infrastructure, the lesson is blunt. Secrets can now leak without a human ever pasting one into the wrong place, and the guardrails around automated coding tools need to catch up.