SGLang RCE Flaw Exploited to Breach NEAR AI, Exposing Data From 18 Agents and ~100 Users

What happened

According to The Hacker News, an attacker exploited a remote code execution (RCE) vulnerability in SGLang, an open-source inference engine, to run code on one of NEAR AI's inference servers last Sunday. NEAR AI said the attacker may also have been able to view requests processed on that server that day, including requests submitted in "confidential mode." That mode is designed to process data inside a trusted execution environment (TEE). The breach affected data tied to 18 agents and about 100 users. No CVE identifier is referenced in the source reporting.

Why it matters

This incident shows that AI inference infrastructure is a direct target for RCE exploitation. A compromise at the inference layer can expose request and response data being processed. The report indicates that even TEE-backed "confidential mode" processing may not have fully shielded data from an attacker with code execution on the host. Any organization running SGLang, or similar open-source inference engines, to serve agents or users should review its exposure, since one compromised inference node can lead to data exposure across many downstream agents and users.

What defenders should watch for

• Inventory where SGLang or other open-source inference engines are deployed in your environment, and check for available patches or security advisories from the SGLang project.
• Review the network exposure of inference servers. RCE flaws in inference-serving software are especially dangerous when these endpoints are internet-facing or broadly reachable.
• Audit logging and monitoring on inference hosts for anomalous process execution, unexpected outbound connections, or unusual file and memory access patterns consistent with post-exploitation activity.
• Reassess assumptions about TEE and confidential-compute protections. Verify whether your deployment's TEE boundary could be bypassed by code execution at the host or orchestration layer, as this report suggests may have occurred.
• If you operate multi-tenant AI agent or inference platforms, consider the blast radius of a single compromised inference server across tenants and agents, and evaluate your isolation boundaries.

Developing story

Details of the exploit mechanics, the specific SGLang vulnerability, and the full scope of impacted data are still emerging, and no CVE has been cited in current reporting. This is net-new intel without a mapped detection at this time. Readers should monitor NEAR AI's and SGLang's own disclosures for updates. Original source: The Hacker News.