Together AI confirmed on Saturday (21/11) a prompt leak from its serverless inference service: an unauthorized party downloaded samples of customer prompts. The official disclosure, signed by CEO and co-founder Vipul Ved Prakash, points to an uncomfortable root cause: while developing the public xorl repository on GitHub, an AI coding agent committed a credential that gave access to the private repository where the samples were stored. The company says its automated scanning detected the exposure and reverted the commit — but during the exposure window, someone obtained the credential and downloaded the samples.
The samples had been collected for routine software testing, with authorization for that use, and were kept in a private Hugging Face repository. They came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, made mainly on four days: June 24 (Wednesday), July 4 (Saturday), July 8 (Wednesday) and July 14 (Tuesday), 2026. The company did not say how many samples were downloaded or what each one contained.
What the downloader did with the samples
According to Together AI, the party did not stop at downloading the samples: it searched the prompts for credentials to other services and tested whether they were valid. Five of the credentials tested were canaries — decoys planted as part of a Dragos security experiment, designed precisely to reveal this kind of unauthorized access. The company is still reviewing whether any real credential was among the samples and whether personal information or other customer data was exposed, and says it will contact affected customers directly.
This episode is the other half of the story Geekiki published today: the Dragos canary research was the same work that led OpenRouter to remove three providers after finding months of prompt theft. Together AI's disclosure shows the problem was not limited to OpenRouter's gateway providers — it also reached Together AI's own testing infrastructure.
What Together AI changed after the incident
After the prompt leak, the company says it worked with an external forensic investigation firm and law enforcement, found no evidence of additional prompt leakage beyond this incident, and has no reason to believe the data leaked further — according to the company, the data has now been deleted. The four changes announced are: scoped or short-lived credentials, and keyless access where supported; checking all pull requests for credentials; zero data retention (ZDR) as the default for new accounts; and updated development processes governing the use of AI coding agents. Customers handling sensitive prompts should enable ZDR in their account settings.
Image credit: official art from Together AI's security disclosure (together.ai).
Sources: Together AI — Security event: Prompt leakage for requests in June and July of 2026; The Hacker News — OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms; Unite.AI — Together AI Discloses Prompt Leak After Coding Agent Exposed Credential.
Transparency note: this article was produced with the assistance of artificial intelligence, cross-checking public posts on X and sources on the internet, and reviewed before publication. Check the original sources to confirm the information.
Por Muse