SGLang RCE Flaw Exploited to Breach NEAR AI Infrastructure
A critical remote code execution vulnerability in SGLang allowed attackers to access NEAR AI's inference servers. The breach potentially exposed data from 18 agents and 100 users.
TL;DR
- Attackers exploited an RCE flaw in SGLang, an open-source inference engine.
- NEAR AI confirmed unauthorized code execution on its inference servers.
- Confidential user requests processed in TEE may have been compromised.
- Data from 18 agents and around 100 users was potentially stolen.
- Organizations using SGLang should audit and patch their deployments immediately.
An attacker leveraged a remote code execution (RCE) vulnerability in SGLang—an open-source inference engine—to gain unauthorized access to one of NEAR AI's inference servers. The incident occurred on a recent Sunday and allowed the threat actor to execute arbitrary code on the compromised server.
According to NEAR AI, the breach also enabled the attacker to view requests being processed that day, including those marked as confidential. These requests were handled within a trusted execution environment (TEE), raising serious concerns about potential exposure of sensitive data despite protective measures.
Vulnerability Details
- The exploited flaw was a remote code execution vulnerability in SGLang.
- SGLang is widely used for running large language model inference workloads.
- The vulnerability allowed full control over affected inference servers.
- No authentication bypass was required; the flaw existed in core processing logic.
Impact and Response
- Approximately 18 agent accounts and 100 user records were accessed.
- Confidential requests processed inside a Trusted Execution Environment (TEE) may have been visible to the attacker.
- NEAR AI has since patched the vulnerable system and initiated a full investigation.
- Affected parties have been notified, and mitigation steps are underway.
Sources
Sources
Security email updates
One digest email when we publish new security articles (TL;DR plus links to read more). Unsubscribe anytime from the message footer. See our Privacy Policy.