Together AI says an AI coding agent committed a credential to a public GitHub repo. That credential opened a private Hugging Face repository full of stored prompt samples. An unauthorized party downloaded the samples, went looking for other services' credentials inside them, and tested what it found. Five of those tested credentials were canaries from a Dragos security experiment. That is the account in Together's own disclosure, published November 21, 2026 by CEO Vipul Ved Prakash, and relayed by The Defiant.
The take: the interesting part is not the leak. It is the chain. An agent wrote a secret into a commit, and a file of prompts became a place to hunt for more secrets. Anyone building agents that hold keys should read it as a checklist.
What Together says happened
Everything below is Together's statement, not independent verification.
- ▸What was stored: prompt samples collected during routine software testing, kept in a private Hugging Face repository.
- ▸Which requests: samples from requests to MiniMax M3, GLM-5.2 and Kimi K2.6, primarily on June 24, July 4, July 8 and July 14, 2026.
- ▸Root cause: an AI coding agent committed a credential while Together was developing its public repository, "xorl". The credential granted access to the private repo.
- ▸Detection: Together's automated scanning flagged the exposed credential, and the commit was reverted.
- ▸Unauthorized activity: during the exposure window, an unauthorized party obtained the credential and downloaded the samples. It searched them for credentials to other services and tested whether they were valid.
- ▸Canaries: five of the tested credentials were canary credentials tied to a Dragos experiment.
- ▸Response: Together says it worked with an external forensic firm and law enforcement, and found no evidence of prompt leakage beyond this incident.
What is still open
The post leaves real gaps, and they matter more than the tidy parts.
- ▸Whether the exposed prompts contained sensitive personal information or other customer data is still under review.
- ▸Together does not say how many prompts, samples or customers are involved. It says it will contact affected customers directly as needed.
- ▸The exact start and end of the exposure window are not given.
- ▸"Primarily" on the dates above means the post does not rule out requests outside that list.
- ▸Five tested credentials were canaries. The post does not say what the other tested credentials were, or whether they were valid. Treat "the intruder only hit canaries" as unsupported. The deck says five were canaries, not that all were.
- ▸The identity and intent of the unauthorized party are not described beyond the actions above.
Why a canary matters here
A canary credential is a fake secret that looks real and phones home when someone tries it. Dragos used them in an experiment, per Together's account. When canaries turn up as tested credentials, that shows someone was mining the samples for secrets. It is one data point about behavior. It says nothing about how much real data was taken.
The pattern is the lesson. People paste keys into prompts, and test logs keep prompts. Both are ordinary habits, and together they turn a stored sample into a credential store. We read only Together's disclosure for this post, so we are not drawing wider conclusions about the Dragos experiment itself.
What a builder takes from it
Onchain agents raise the stakes, because the credential in question can be a signing key, an RPC key or a facilitator token. Together's own remediation list doubles as a build checklist:
1. Scoped or short-lived credentials, keyless where supported. A leaked token that expires or can do one thing is a smaller incident. 2. Check every pull request for credentials. Scanning caught this one, but only after a coding agent had already committed the secret. 3. Zero data retention as the default. Together is making ZDR the default for new accounts. Its post does not say existing accounts changed, and it advises customers handling sensitive prompt data to enable ZDR in their settings. 4. Rules for coding agents. Together says it is updating its development processes for AI coding agents. If an agent can commit, it needs the same secret-handling rules as a person, enforced by tooling rather than hope.
If your agent can move funds, "we scan commits" is a floor. Keep spend keys out of prompts and test fixtures entirely, and keep hot keys gas-only.
What to watch
- ▸Whether Together publishes the exposure window, sample counts and an answer on sensitive data.
- ▸Whether it says the exposed credential was revoked. The post describes process changes but does not state it.
- ▸Whether other inference providers publish retention and credential-handling details in response.
- ▸Whether ZDR defaults extend to existing accounts.
Building an agent that touches keys? Describe it at satohub.ai/build and Sato maps the wallet, payment and key-handling stack, so the secret-handling question comes before the first commit.