Together AI Says Coding Agent Leaked Credential, Exposing Customer Prompts
Together AI says an AI coding agent leaked a repo credential during xorl development, letting an intruder download stored customer prompt samples and test embedded secrets for validity.
By Nathan Brooks
3 min read

What's News
- Together AI disclosed the prompt leak on November 21, 2026, tied to prompts sampled on June 24, July 4, July 8 and July 14, 2026.
- An AI coding agent exposed a credential to a private Hugging Face repository during development of the public xorl GitHub repository.
- The intruder searched leaked prompts for other credentials and tested validity; five tested credentials were Dragos canary credentials.
- Zero data retention (ZDR) is now the default for new Together AI accounts.
- Affected prompts involved MiniMax M3, GLM-5.2 and Kimi K2.6 model requests.
Together AI disclosed on November 21, 2026 that an unauthorized party downloaded stored samples of customer prompts from its serverless inference service — after an AI coding agent leaked the credential that unlocked them.
The prompts came from requests customers had authorized for routine software testing. Together AI had stored the samples in a private Hugging Face repository. They covered requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, primarily on June 24, July 4, July 8 and July 14, 2026.
How did the credential get out?
The leak traces back to Together AI's development of its public xorl GitHub repository. During that work, a coding agent committed a credential granting access to the private Hugging Face repository.
Together AI's automated scanning caught the exposure, and the company reverted the commit. That wasn't enough. During the exposure window, an unauthorized party grabbed the credential and used it to download the stored prompt samples. The sequence makes an uncomfortable point for anyone running AI-assisted development: detecting and reverting a commit does not invalidate a secret that someone already copied.
The intruder didn't stop at reading. After the download, the party searched the samples for credentials to other services and tested whether they were valid. Five of the tested credentials were canary credentials tied to a Dragos security experiment. So the risk went beyond prompt text disclosure — it extended to possible reuse of secrets customers had embedded in their prompts.
What does this change for customers?
Impact depends entirely on what was inside the sampled prompts. Together AI's review is ongoing to determine whether sensitive personal information or other customer data was exposed. The company said it will contact affected customers directly as needed with findings relevant to them.
Together AI worked with an external forensic investigation firm and law enforcement on the response. The company said its review has not identified evidence of additional prompt leakage beyond this incident, that it has no reason to believe the data was leaked further, and that it understands the data has now been deleted.
For practitioners, the practical checklist is short: audit what secrets live inside your prompts, assume stored samples are a standing attack surface, and figure out whether your prompts from those four July dates contained credentials at all.
What is Together AI changing?
The company's remediation package covers both halves of the failure. On the credential side:
- Scoped or short-lived credentials and keyless access where supported, designed to reduce the potential impact of any future exposure.
- Credential checks on all pull requests.
- Updated development processes governing the use of AI coding agents.
On the retention side, zero data retention — ZDR — is now the default setting for new accounts, limiting prompt exposure risk by not storing prompt samples at all. Together AI recommends that customers handling sensitive prompt data enable ZDR in their account settings. The company has set up [email protected] for questions about the incident or its potential impact on an organization.
The incident links two problems that usually live in separate budgets: credential safety in AI-assisted development, and the retention of customer prompts for testing. Together AI has now documented controls for both. But the ongoing content review remains the decisive step — it will determine what the exposure actually means for customers and which notifications become necessary.
Source: unite.ai
More from Nathan Brooks
Show full bio
News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.
2,611 articles