NEWS

SGLang RCE Exploit Breaches NEAR AI, Exposing 18 Agents and About 100 Users

An attacker ran code on a NEAR AI inference server via an SGLang RCE flaw, exposing data from 18 agents, about 100 Private Chat users and two API keys.

Dylan H.

News Desk

November 10, 2026
3 min read
SGLang RCE Exploit Breaches NEAR AI, Exposing 18 Agents and About 100 Users

Overview

NEAR AI disclosed that an attacker exploited a remote code execution flaw in SGLang, an open-source LLM inference engine, and ran code on one of its inference servers on Sunday. The company has not named a CVE, the affected SGLang versions, or whether upstream has shipped a fix. The incident was reported by The Hacker News on 2026-11-10.

Notably, the server ran inside a trusted execution environment (TEE), which did not protect against a software vulnerability exploited inside it.

What Was Exposed

  • About 100 Private Chat users: email addresses and names, and for some, plan and balance information. NEAR AI found no evidence that stored conversations or uploaded files were accessed.
  • 18 AI agents on an older NEAR AI Agents deployment outside the TEE: chat history, settings and logs.
  • Two customers' NEAR AI Cloud API keys, exposed in logs. Both were revoked.
  • Possible live-traffic visibility: the attacker could have seen requests processed on that server that day, including confidential-mode requests. NEAR AI has not said whether those users were contacted.

Brave's Leo assistant runs some chats in a NEAR AI-operated TEE. NEAR AI has not said whether the affected server handled any of them.

NEAR AI's Response

  • Revoked all active sessions and admin credentials and replaced agent tokens
  • Took Private Chat offline (it was already deprecated)
  • Patched SGLang on all inference servers and rotated remaining credentials
  • Completed a forensic review and restored NEAR AI Cloud access
  • Reported the flaw to the SGLang project

Affected users were contacted directly. Others were told no action is needed.

Why It Matters

SGLang has been a repeat target. Labs previously covered SGLang CVE-2026-5760, a CVSS 9.8 RCE via malicious GGUF model files. Inference engines process untrusted input at high privilege and often sit next to API keys and customer prompts, so they deserve the same patching urgency as any internet-facing service. TEEs protect data from the host, not from bugs in the workload itself.

Recommendations

  1. Inventory SGLang deployments and upgrade to the latest release.
  2. Keep inference servers off the public internet, or put them behind authenticated gateways.
  3. Never log API keys or bearer tokens, and scrub secrets from logs.
  4. Rotate agent tokens and API keys after any inference-host compromise.
  5. Be cautious of unsolicited messages about accounts or balances if you used NEAR AI Private Chat.

Sources