Together AI discloses prompt leakage: AI coding agent committed exposed credential

togethercompute · x · 2026-11-22

Together AI disclosed a security incident: prompts sent to its serverless API in June–July 2026 (requests to MiniMax M3, GLM-5.2, Kimi K2.6) were downloaded by an unauthorized party. Root cause: while developing the public xorl GitHub repo, an AI coding agent committed a credential to a private Hugging Face repo storing prompt samples. Automated scanning caught it and the commit was reverted, but during the exposure window the credential was used to download samples; the attacker then searched them for credentials to other services. The company is reviewing affected data and notifying customers.

Original post →

More from Infra

Infra channel →