NEAR AI Discloses Breach After Attackers Exploit Unpatched SGLang Vulnerability

NEAR AI contained an unauthorized breach of its infrastructure after attackers exploited an unpatched remote code execution vulnerability in SGLang, the open-source inference engine used across its servers ². The attacker executed arbitrary code on a single inference node, pivoted into an internal system supporting deprecated Private Chat infrastructure, and downloaded account metadata for roughly 100 users, including email addresses, names, and selective subscription and balance details ².

The incident also compromised 18 agents operating on a legacy deployment lacking Trusted Execution Environment (TEE) protections, exposing chat histories, configuration settings, and execution logs ². On Nov. 8, the intruder intercepted user requests actively handled by the compromised server, capturing requests processed even under confidential mode ². In addition, NEAR AI identified that API keys belonging to two Cloud customers were exposed in server logs; the platform has since revoked these credentials ².

Unpatched Vulnerabilities Threaten AI Inference Infrastructure

The attack underscores systemic supply-chain risks across open-source AI infrastructure, similar to incidents highlighted when Chainalysis disclosed a security breach involving agentic AI tools. According to a vulnerability advisory published by CERT/CC, researchers identified six distinct security flaws in the SGLang framework ¹. These issues span remote code execution through deserialization bypasses, server-side request forgery, credential exposure via the /server_info endpoint, and model weight extraction via distributed communication protocols ¹. CERT/CC reported that maintainers have not released official patches and coordination attempts have stalled, leaving default configurations vulnerable unless manually restricted ¹.

NEAR AI confirmed it deployed internal fixes across its inference instances, rotated administrative credentials, revoked agent tokens, and notified affected parties ². The firm observed that while confidential computing protects workloads from host tampering, it cannot defend against vulnerabilities existing inside the application software itself ². At the time of writing, native token $NEAR traded down 1.3% over the prior 24 hours at $3.58.