NEAR AI’s security disclosure details how an attacker exploited a remote code execution flaw in SGLang, an open-source inference engine the company uses for model serving, on November 8. The breach exposed account details (email, name, subscription balances) for roughly 100 users and chat logs from 18 AI agents. Two Cloud API keys found in logs were also compromised. NEAR AI contained the intrusion by November 9, revoked credentials, and confirmed no stored Private Chat conversations or uploaded files were accessed.
The troubling part is the disclosure’s vagueness. NEAR AI reported the flaw to SGLang developers but named no CVE or affected versions publicly. Multiple critical SGLang RCEs exist in the wild (CVE-2026-5760, CVE-2026-3059, CVE-2026-86793), and without clarity on which one caused the breach, other organizations can’t patch with confidence. Meanwhile, the exposed data sets the stage for downstream attacks: user email addresses and names feed password-spray and social-engineering campaigns. The breach also reveals infrastructure risk stratification at NEAR AI. The exposed agent logs came from older, non-TEE deployments, while newer systems sit behind better isolation.
Sources: NEAR AI Security Incident Report: SGLang Vulnerability
Coverage: SGLang RCE Exploit Breaches NEAR AI, Steals Data From 18 Agents and About 100 Users
Related on this blog: [AUTO] AI-Driven Patch Flood: Signs of Slowing, Not Peak • [AUTO] Anthropic’s automated vulnerability scanner reaches 85% accuracy. Can it scale? • [AUTO] Vulnerability discovery hits the patch ceiling