The commit that opened the store
Together AI's notice is short, and its central fact is stated without softening. During development of the company's public GitHub repository, xorl, an AI coding agent committed a credential. That credential provided access to a private Hugging Face repository where Together stored samples of customer prompts, collected as part of routine software testing.
Automated scanning did the job it is there for. It detected the exposed credential, and Together reverted the commit. The sequence still failed where it mattered: during the exposure window, an unauthorized party obtained the credential and used it to download the stored samples. A revert removes a secret from the current tree. It does not call back a copy already taken. The notice does not say whether the download came before or after the detection, and it gives no length for the window; what it establishes is that detection and a revert were not enough to prevent the download.
The samples were not live traffic intercepted in transit. They were a stored set: prompts submitted to Together's serverless inference service in June and July 2026, from requests to three models, MiniMax M3, GLM-5.2 and Kimi K2.6, concentrated on June 24, July 4, July 8 and July 14. Together says the samples were authorized for the testing use. That covers why the company held them. It does not answer how a credential able to read the whole set came to be committed from a development process working on a public repository, and the notice does not explain how the agent obtained it.
The intruder read the prompts for secrets
What the unauthorized party did with the download is the second fact readers should keep. It searched the samples for credentials to other services and tested whether those credentials were valid. Testing is not confirmation: the notice does not say any of the found credentials worked.
That is a predictable use of a prompt corpus. Prompts routinely contain pasted configuration, keys, internal URLs and fragments of code that users sent to a model without treating the prompt itself as a stored document. Five of the credentials the party tested turned out to be canary credentials associated with a Dragos security experiment: planted values designed to register when someone uses them. Together states that connection in its own notice. ALFW covered a separate disclosure this week, OpenRouter's removal of a zero-retention label from MiniMax endpoints after canary URLs were opened; no source establishes a connection between that incident and this one, and the mechanisms differ. OpenRouter's case concerns prompts routed to a model provider. Together's concerns a sample store the company kept itself, opened by a credential its own agent committed.
Together says it has worked with an external forensic investigation firm and law enforcement, and that its review has not identified evidence of additional prompt leakage beyond this incident. Its review of what the samples actually contain is still running, including whether sensitive personal information or other customer data was exposed, and it says affected customers will be contacted directly as needed.
What the notice leaves blank
Three numbers a reader needs are absent. The notice gives no date for the credential exposure and no length for the exposure window, so there is no way to judge how long the private repository was reachable with a public credential. It gives no count of prompts or customers in the sample set; "a subset" and four main collection dates are the only sizing provided. And it does not say whether the downloaded data has been deleted or recovered. Secondary coverage has added wording on that point that the notice itself, read first-hand, does not contain. Until Together states it, deletion is unconfirmed.
The remedy list is more specific. Together says it now uses scoped or short-lived credentials, and keyless access where supported; it checks all pull requests for credentials; it has updated the development processes governing its use of AI coding agents; and it has made zero data retention the default setting for new accounts.
Read that last item with its qualifier. The default changed for accounts created after the change. Existing customers handling sensitive prompt data are told, in the same notice, to enable ZDR themselves in their account settings. For a current customer, the protection is still a setting to go and switch on, dated from a disclosure published on November 21, 2026.
What to take from it
The practical lesson is not that a coding agent made a novel mistake. Committing a credential to a public repository is one of the oldest errors in software. The lesson is about reach, stated generally: an automated contributor that can commit to a public repository should not be able to reach credentials that read private customer data, because the review and scanning layers behind it are detection, not prevention. Together's own account shows the gap: scanning worked, the revert worked, and the download happened inside the same window.
For Together customers, the notice supports one direct check and one honest limit. The check: whether ZDR is actually enabled on the account rather than assumed from the new default. The limit: the four named dates and three named models describe where the samples mainly came from, but they cannot tell any individual customer whether a particular request was in the sample set. Only Together's direct notification can do that.
Sources
- [1] Together AI, "Security event: Prompt leakage for requests in June and July of 2026," published November 21, 2026, updated November 22, 2026Read source
- [2] The Hacker News, "OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms," November 2026 (secondary roundup placing the Together notice alongside the wider canary research; not used for any fact stated as Together's own)Read source