Incident
Published 2026-11-19 New
Verified 2026-11-22

Together AI says an unauthorised party downloaded stored customer prompt samples after an AI coding agent exposed a credential; OpenRouter says it off-boarded three inference providers after researchers' findings

Together AI said in a notice dated 21 November 2026 that an unauthorised party downloaded stored prompt samples from its serverless inference service after a credential for a private repository was exposed in a public code repository (source: Together AI).

Together AI said the samples, collected for software testing and kept in a private Hugging Face repository, came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, primarily on 24 June and 4, 8 and 14 July 2026; that an AI coding agent committed the credential while developing its public GitHub repository xorl; and that the party searched the samples for credentials to other services and tested whether they were valid, five of which were canary credentials from a Dragos security experiment (source: Together AI).

Together AI said it is working with an external forensic firm and law enforcement, has not identified evidence of additional prompt leakage, is still reviewing whether sensitive personal information or other customer data was exposed, will contact affected customers directly as needed, and has made zero data retention (ZDR) the default for new accounts (source: Together AI).

In research published on 19 November 2026, Dragos said it sent about two million prompts carrying unique URLs and canary credentials to AI services from July to November 2026 and found attackers stealing data from three hosting providers, none of which knew of the breach until Dragos notified them; Dragos did not name the providers (source: Dragos).

Dragos said stolen AWS test keys were used to list S3 buckets, search Secrets Manager and Parameter Store, and try to enable AWS Bedrock models and add IAM users, and that some leaked requests had been marked ZDR (source: Dragos).

OpenRouter said on 22 November 2026 that researchers' findings did not show any breach of OpenRouter, that it has off-boarded three providers based on its investigations, removed 125 endpoints from ZDR eligibility (including some as a precaution) and notified all potentially affected customers (source: OpenRouter).

Product
Together AI serverless inference (stored prompt samples); AI inference providers and routers
Versions
n/a — incident
Exploited in Australia?
unknown
Patch to
Together AI says customers handling sensitive prompt data should enable zero data retention (ZDR) in their account settings and can contact security@together.ai with questions.

breaches ai