An attacker used a remote code execution flaw

On Nov 8, an attacker used a remote code execution flaw in SGLang, the open-source AI inference engine, to run code on one NEAR AI inference server and then move into internal systems, per NEAR AI.

Data of about 100 users was accessed: names, emails, and for some subscription and balance info. 18 agents from legacy infrastructure were retrieved, including chat history, settings and logs.

API keys of two NEAR AI Cloud customers were exposed in logs and have been revoked. Requests processed on that server on Nov 8, including confidential inference requests, could potentially have been read.

NEAR AI found no indication stored Private Chat conversations or uploaded files were accessed. It revoked sessions and admin credentials, replaced agent tokens, took Private Chat offline, fixed SGLang on all servers, rotated remaining credentials, completed forensic review, and notified affected users from [email protected].

The key point: confidential computing protects against the operator and host, but not against vulnerable software inside the environment, so code running inside the server could see even confidential-mode requests.

NEAR Protocol, NEAR Intents and other blockchain infrastructure were not affected, and no direct financial loss has been reported.