Together AI published a security notice on November 21, signed by CEO and co-founder Vipul Ved Prakash. It says that some of the prompts sent to its serverless inference service in June and July of this year were downloaded by an unauthorized party. The source was a single access credential: while the company was using an AI coding agent to develop a public code repository, the agent committed the credential along with the code.

How one credential got out
According to the notice, Together routinely samples a portion of user requests for software testing. Those samples were stored in a private Hugging Face repository, a use the company describes as authorized. The samples came from requests sent to MiniMax M3, GLM-5.2 and Kimi K2.6, concentrated on June 24, July 4, July 8 and July 14.
The failure happened elsewhere. Together has a public GitHub repository called xorl, and during development an AI coding agent committed code that included a credential granting access to that private repository. Automated scanning caught it, and Together pulled the commit. But during the window in which the credential was exposed, someone had already obtained it and downloaded the stored prompt samples.
The intruder then did two things: combed through the samples for credentials to other services, and tested them one by one to see whether they worked. What the downloader was after were the keys, passwords and tokens that users had written into their prompts.
Five decoy credentials
One detail in the notice stands out: among the credentials that were tested, five were “canary” credentials belonging to an experiment by the security firm Dragos.
The experiment was recently reported by Ars Technica. Dragos chief research officer HD Moore wrote nearly 2 million fake keys and fake links carrying tracking markers into prompts, sending each one to only a single provider, then watched to see which ones were touched. His finding was that requests at no fewer than three inference hosts had been copied, and that the decoy AWS keys were used tens of thousands of times. The reminder email OpenRouter sent to users about MiniMax models also cites this research.
Whether Together is one of the three hosts Moore refers to is not stated in the notice, and it cannot be confirmed from public information. The only thing that lines up is that Moore’s decoys appeared in the downloaded Together samples and that someone tried them out.
We have no reason to believe the unauthorized party has leaked this data further.
Together also said it believes the data has been deleted. The notice does not say who verified the deletion or how.
The numbers it leaves out
The notice gives dates and model names but not the scale. It does not say how many prompts were affected, how many customers were involved, how long the credential sat in the public repository, or on what day it was pulled. Together says it is still reviewing whether the samples contain sensitive personal information, and customers who need to be notified will be contacted individually. The company has brought in an outside forensics firm and reported the incident to law enforcement.
Its remediation list has four items:
- Switch credentials to scoped or short-lived ones, and avoid keys altogether where possible;
- Run a credential check on every merge request;
- Enable zero data retention by default on new accounts;
- Update the development workflow for using AI coding agents.
Existing accounts were not switched over automatically. The notice only suggests that customers handling sensitive data turn the setting on themselves in their settings.
Reading it alongside the OpenRouter email
Within a few days, two inference providers have had to answer to users over the same research, and their handling differs considerably.
OpenRouter emailed only some users and published no public page; it pointed responsibility at upstream MiniMax and at the providers that relayed the requests, and its action was to remove the zero-data-retention label from the relevant endpoints. Together published a public blog post signed by its CEO, acknowledged that the problem lay in its own development process, and changed the default setting for new accounts.
The two incidents also differ in nature. In the OpenRouter case, a link inside a prompt was visited dozens of hours after the request was sent, and the visitor’s identity is still unknown. In Together’s case there is a clear point of entry: a credential an agent wrote into a public repository, plus a downloader specifically hunting for keys.
All three models behind the downloaded samples come from Chinese companies. Teams in China building products for overseas markets often call Chinese models through overseas hosts such as Together. Anyone who used these three models in June or July and had pasted database connection strings or API tokens into prompts would be wise to rotate those credentials now rather than wait for an individual notice from Together.
There is also the coding agent itself. Committing a key to a public repository is a mistake human engineers make often, and the scanner did catch it this time. Together has not disclosed how long the window lasted before the catch, and the notice does not name which coding agent was involved.
Sources: Together AI official security notice, CocoLoop, Ars Technica; Together’s notice verifies the affected models, the four days on which the samples cluster and the four remediation items, while Ars Technica verifies the scale of the Dragos experiment’s decoys and the number of compromised hosts.