If You Sent Together AI Prompts in June or July, Check Your Settings
The company says an AI coding agent exposed a credential that unlocked a private repository of prompt samples. It has not disclosed how many customers were affected, and its review of the contents continues.
Together AI has disclosed that an unauthorized party downloaded stored samples of customers' AI prompts after obtaining an exposed credential, putting the contents of those requests beyond the company's control. [1]
The company published its notice on November 21, 2026, saying the incident involved a subset of requests submitted to its serverless inference service in June and July. That service lets customers send requests to hosted AI models without running the computing infrastructure themselves. [1]
The samples came from requests to three models, MiniMax M3, GLM-5.2 and Kimi K2.6, primarily on June 24, July 4, July 8 and July 14. Together says the prompts were authorized for routine software testing and stored in a private repository on Hugging Face, a platform for sharing models, datasets and software. [1]
A prompt leak exposes what a customer sends into an AI service. That can be much more sensitive than the answer that comes back.
Cybersecurity company Dragos says prompts can contain confidential source code, internal communications, strategic plans, customer information and credentials. Those are examples of the risk, not established contents of Together's downloaded samples. [2]
Together has not disclosed the number of affected requests, customers or organizations, the precise exposure and download timeline, or what sensitive information the samples contained. Its content review is continuing, including whether personal information or other customer data was exposed. Whether real credentials were exposed or successfully used, and whether requests covered by zero data retention were included, has not been established. [1]
A private repository, a public credential
According to Together, the route into the stored prompts began while the company was developing xorl, its public code repository on GitHub. An AI coding agent committed a credential that granted access to the private Hugging Face repository. [1]
A credential is a digital key. A repository can remain private while a publicly exposed key lets someone open it. Here, the prompts themselves did not have to be published in the code repository for the protection around them to fail.
Together says its automated scanning detected the credential and the commit was reverted. But an unauthorized party obtained the credential during the exposure window and used it to download the samples. The company's account describes detection that came too late to prevent access. [1]
The intruder then searched the downloaded prompts for credentials to other services and tested whether they worked, according to Together. Five of the credentials tested were canaries associated with a Dragos security experiment. [1]
Together says it worked with an external forensic investigation firm and law enforcement. It says its review has found no evidence of prompt leakage beyond this incident, has no reason to believe the intruder distributed the data further, and understands that the data has been deleted. Those are Together's assessments, not independently verified findings. [1]
Deleting a downloaded copy would be welcome. It would not undo the access or the subsequent search through the prompts for other keys. The immediate task for customers is to learn what information left the service and whether anything in it requires action.
The planted keys that revealed access
Dragos published its account of the experiment on November 19. It says it seeded nearly two million prompts with unique URLs and canary credentials between July and November 2026 to detect whether information from prompts was accessed. [2]
A canary credential is a planted key used as a warning signal. Its value lies in revealing that someone found and tried it. Together's account of five tested canaries therefore describes more than a download, it describes someone examining the contents for usable access to other systems. [1]
Dragos describes evidence of exposure at three or more AI providers, but does not name them in its public account. Its broader findings do not independently identify Together as one of those providers. The link to this incident comes from Together's own disclosure. [1] [2]
The experiment also illustrates why the path a prompt takes matters. According to Dragos, AI requests may pass through multiple providers or routers, services that receive a request and forward it to the model provider that handles it. Each can see prompt and response content. A customer's choice of model can involve a chain of services handling the same material. [2]
In Together's account, the exposed material was a stored testing sample. That puts the handling of prompts after a request alongside the security of the request itself. Every retained copy creates another place that needs protection.
Check the retention setting
Together says it has strengthened its controls by using credentials with narrower permissions or shorter lifetimes, adopting access without stored keys where supported, checking all proposed code changes for credentials, and changing its development processes for AI coding agents. [1]
These measures address the failure the company describes: an agent placed a powerful key into public code. Reviewing agent-generated changes before publication is a basic security boundary, especially when those changes can expose customer data.
For customers, Together recommends enabling zero data retention, or ZDR, when handling sensitive prompts. Its documentation defines that setting as preventing storage of prompts and model outputs or their use for secondary purposes. Some request metadata is still retained. [1] [3]
The setting is prospective. It applies only to requests made after it is enabled, so switching it on now does not remove previously stored prompt samples. Disabling prompt storage also disables passthrough models, which forward requests to third-party providers. [3] [4]
There is an apparent mismatch in Together's instructions. The incident notice says ZDR is now the default for new accounts. Its documentation says ZDR is not enabled by default, while its privacy page says prompt storage is on by default. Customers handling sensitive material should check their organization's actual setting rather than assume it from either description. [1] [3] [4]
Every edition in brief, three times a day, on our Telegram channel, on Bluesky and on Threads.
Spotted an error? Tell the editors
- Security event: Prompt leakage for requests in June and July of 2026 Hacker News
- AI Supply Chain Security: What Canary Tokens Found | Dragos - Dragos dragos.com
- Zero data retention - Together AI docs docs.together.ai
- Privacy and security - Together AI docs docs.together.ai




