SGLang Flaw Exploited, Breaches NEAR AI Systems, Data Stolen

An attacker exploited a flaw in the SGLang inference engine to run code on a NEAR AI inference server and steal data from 18 AI agents plus sensitive account details of around 100 users, according to a breach disclosure. The compromise occurred even though the server was running in a secure, confidential mode within a trusted execution environment. The incident shows that confidential-computing protections did not stop exploitation of the SGLang inference engine flaw.