An attacker used an SGLang fault to get data from 18 NEAR AI agents

An attacker used a fault in SGLang, an open-source engine for AI models, to operate code on a NEAR AI server last Sunday. NEAR AI says the attacker got account data of about 100 users and data from 18 AI agents. The company closed the access and revoked two keys of customers that were in logs.

How it works

The server operated SGLang in a trusted execution environment, an area made to keep queries private from NEAR AI. NEAR AI says a fault in software in the environment is not prevented by it. It is possible that the attacker saw queries in confidential mode that day.

What was taken

• About 100 users of Private Chat: email addresses, names and, for some, plan and balance data.
• 18 agents in a deployment not in the environment: chat history, settings and logs.
• Two keys of customers, in logs.

NEAR AI found no sign that the attacker read the chats or files that users put in Private Chat.

What is not known

NEAR AI gave no CVE number, no SGLang versions and no information on a repair from the makers of SGLang. Brave says that its Leo assistant uses a NEAR AI environment. NEAR AI did not tell if some of those chats went to that server.

Insight: Logs and old systems are the weak point

In October, Wiz reported breaches at 9 or more AI developer platforms, where attackers went for keys. Here, two keys of customers were in NEAR AI logs, and the 18 agents were not in the environment. This reads as a sign that attackers go for logs and systems not in the environment. Your risk comes from the vendor part that is not in the environment.