NEAR AI breach via SGLang RCE exposes about 100 users and 18 agents

NEAR AI says an SGLang remote code execution flaw exposed about 100 users and 18 agents; keys were revoked and no CVE was named.

NEAR AI says an attacker exploited a remote code execution flaw in SGLang, an open-source model-serving engine, on one inference server and then reached other infrastructure. The attacker could view that day’s requests, including confidential-mode traffic inside a TEE, and later reached deprecated Private Chat and an older agent deployment outside the TEE. Details for about 100 users and data from 18 agents were downloaded, including two customers’ Cloud API keys. Private Chat exposure covered emails, names and, for some users, plan and balance data; NEAR AI says stored conversations and uploaded files were not found accessed. Agent data included chat history, settings and logs. NEAR AI revoked the keys and sessions, patched SGLang on its servers, restored NEAR AI Cloud and reported the flaw upstream, without naming a CVE, versions or an upstream fix.