• Sources: discussion
  • Summary: The Hacker News reports that an attacker exploited an SGLang remote code execution flaw on one of NEAR AI's inference servers, could see that day's requests including confidential mode requests processed in a trusted execution environment, and then reached 18 agents on an older deployment and account details of about 100 Private Chat users. Two customers' API keys were exposed in logs and revoked. NEAR AI has deployed a fix on all its inference servers and reported the flaw to SGLang, but no CVE, affected versions or upstream release are public.
  • Why it matters: Anyone exposing a self-hosted SGLang server should restrict network access to it until the project publishes a fix and affected versions.
  • Background: SGLang is an open-source LLM serving engine used as an alternative to vLLM. A TEE encrypts and isolates a workload from the host operator, but code running inside the enclave, here the inference engine itself, is still in the trust boundary, which is the point NEAR AI makes. Brave said in October that its Leo assistant runs some chats in a TEE operated by NEAR AI.
  • Follow-up: An SGLang advisory with a CVE ID and affected versions.