AI is going just great.
ConcerningModeratetogether-ai

Together AI coding agent committed a credential to a public repo, exposing stored customer prompts

Logged · Source: news.sulat.com ↗

We had a security event that resulted in leakage of a subset of prompts sent to our serverless APIs in June and July of 2026.

An AI coding agent committed an internal credential to Together AI's public xorl GitHub repository, giving an unauthorised party access to a private Hugging Face repository holding prompt samples from the company's serverless inference service. The stolen samples came from requests to MiniMax M3, GLM-5.2, and Kimi K2.6, primarily on 24 June and 4, 8, and 14 July 2026. CEO Vipul Ved Prakash posted the security notice on 21 November; Together says its review of the affected prompts is still running to determine whether sensitive personal information was exposed.

The incident intersects with a separate Dragos experiment: after downloading the samples, the unauthorised party tested credentials found inside them for validity. Five of those credentials were canary tokens planted by Dragos as part of a security study. Dragos published results on 19 November showing that canary-seeded prompts sent to AI services between July and November 2026 showed evidence of exposure at three unnamed providers. Together's notice doesn't confirm it is one of them. Together has since made zero data retention the default for new accounts.

Sources

Archived copy of the source ↗

ShareXBlueskyThreads