Skip to content
Rush Commerce
AI & Automation3 min read

Together AI prompt leak: an AI coding agent committed the key

Together AI says an AI coding agent pushed a credential to GitHub, exposing stored prompt samples from June and July. ZDR is now default for new accounts only.

The leak did not start with a hacker. It started with a commit. In a security notice dated November 21, Together AI CEO Vipul Ved Prakash says an AI coding agent committed a credential to a public GitHub repository. That credential opened a private repository of stored prompt samples, and an unauthorized party downloaded them. This is the AI coding agent secrets problem in one paragraph, and it hit an inference provider that sells to developers.

What actually happened

Per Together's notice:

  • The agent pushed the key. While working on Together's public repo, xorl, "an AI coding agent committed a credential" that gave access to a private Hugging Face repository.
  • What was in that repo. Prompt samples from Together's serverless inference service, kept "as part of routine software testing." They covered requests to MiniMax M3, GLM-5.2, and Kimi K2.6, mainly from June 24, July 4, July 8, and July 14, 2026.
  • What the attacker did. They searched the samples for credentials to other services and tested whether those worked. Five were canary credentials from the Dragos experiment. That is how this surfaced.
  • Detection. Together's automated scanning caught the exposed credential and the commit was reverted. The notice does not say how long the key was public.
  • Still open. Together is working with a forensic firm and law enforcement. Its review of whether personal or customer data was exposed is ongoing, and it will contact affected customers directly.

Together's fixes: scoped or short-lived credentials, credential checks on every pull request, new rules for how its engineers use AI coding agents, and ZDR as the default for new accounts. For everyone else, the notice says customers handling sensitive prompt data "should enable ZDR in their account settings."

Why it matters for your business

Two lessons, one for each side of the API.

As a customer: your prompts were a test fixture. A provider kept real traffic samples for testing, and that copy is what leaked. ZDR is now the default for new Together accounts, not old ones. If you have an existing account, go turn it on.

As a builder: your team probably runs coding agents with write access to repos right now. The agent does not know a key is a key. It sees a string that makes the test pass. Together had scanning in place and still had a public window.

What we would put in place:

  • Secret scanning that blocks the push, not one that alerts after it lands.
  • No long-lived keys on dev machines where agents run. Short-lived tokens limit the blast radius.
  • Agent commits get human review before they reach a public branch.
  • No production data in test fixtures. Synthetic prompts only.

Key takeaways

  • An AI coding agent committed a credential to Together AI's public xorl repo
  • The key exposed stored prompt samples for MiniMax M3, GLM-5.2, and Kimi K2.6 from June and July 2026
  • Dragos canary credentials in those prompts revealed the theft
  • ZDR is now default for new Together accounts only; existing customers must turn it on
  • Block secrets at push time and review every agent commit before it goes public

Running coding agents with repo access? We set up push-blocking secret scanning, short-lived credentials, and review gates so your agents ship code, not keys. See what we build or talk to us about your agent setup.

Sources: Together AI: Security event: Prompt leakage for requests in June and July of 2026, OpenRouter: Data Processing Incidents on Inference Providers.

  • #together-ai
  • #ai-coding-agents
  • #secrets-management
  • #prompt-privacy
  • #zero-data-retention
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.