NEAR AI breach exposes user data and agent histories after SGLang flaw
The intrusion reached data from about 100 Private Chat accounts and 18 agents, and shows that a confidential-computing enclave does not shield bugs in the software inside it.
An attacker abused a remote code execution weakness in SGLang, an open-source engine for running AI models, to gain control of a NEAR AI inference server on Sunday. Requests handled on that machine that day, including those sent in the company's confidential mode, could have been visible to the intruder, even though that mode is meant to keep prompts hidden inside a trusted execution environment. Because the flaw lived in the software running within that environment, the protection did not cover it. The intruder then moved into a Private Chat system the company had already retired and pulled account details for roughly 100 users, while separately extracting chat histories, settings and logs from 18 agents on an older deployment outside the protected environment. NEAR AI has since closed the access, patched SGLang across its inference fleet, rotated credentials and revoked two exposed cloud API keys, though it has not published a CVE number or the affected SGLang versions.
Why it matters
For people who rely on hardware-isolated AI services, the incident is a reminder that a trusted execution environment protects against outside access but not against flaws in the code it runs. Users of the affected products had their account details and conversation records taken, and the case shows that retired systems and deployments outside the protected environment can remain a route to sensitive data. The company has closed the access and patched the software, but has not yet published the vulnerability identifier or the affected versions, so outside parties cannot yet check how widely the flaw applied.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.