NEAR AI Contains Security Breach After SGLang Vulnerability Exploitation

NEAR AI runs cloud infrastructure for confidential AI workloads built on SGLang, the open-source inference engine many providers use to serve large language models. The company had marketed this setup as hardware-backed and privacy-preserving before pausing service and then publishing, on November 9, an incident report describing how a remote code execution flaw in SGLang let an attacker move from a single inference server into its internal infrastructure.

About 100 users now have account details that were downloaded by the attacker, and NEAR AI must notify them and explain what was taken. Eighteen legacy agents exposed chat history and logs, data their operators may need to review for leaked content. Two customers had API keys exposed and already revoked, so they must rotate credentials and re-authenticate any integrations before resuming normal use.

The breach exposes the limits of hardware-backed privacy claims: confidential AI requests that were supposed to stay sealed inside secure infrastructure became visible once an attacker reached the inference layer through a software flaw, not a hardware one. That gap between marketed confidentiality and what a single unpatched dependency can undo is the real story here, and it is a fair question whether 'confidential computing' labels oversell guarantees that one vulnerable open-source component can break.

Watch whether NEAR AI discloses which parts of its internal infrastructure the attacker actually reached beyond the one compromised inference server, a detail the November 9 report left open. Also watch whether other operators running SGLang confirm they have patched the same remote code execution flaw, and whether NEAR AI commits to an independent audit of its hardware-backed privacy claims following the cloud restoration.

NEAR ProtocolNEAR it is the asset of the company or protocol in the event

NEAR Protocol fell 1.1% in the 4 hours after the first headline, after rising 1.4% in the 2 hours before.

48 prices, one every 30 minutes · The move was larger before the first headline than after: an observation, not an explanation. The market also moves on things nobody publishes.