RCE in SGLang let attacker see requests in NEAR AI's confidential inference mode
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
An enclave protects against the server operator but not against a hole in the code inside it — the promise of "confidential inference" can no longer be made without auditing the entire stack under the TEE.
According to a NEAR AI report published on November 9, an attacker exploited a remote code execution (RCE) vulnerability in the open-source SGLang inference engine to gain code execution on an inference server running inside a trusted execution environment (TEE). Because SGLang ran inside the enclave, on November 8 the attacker could see the requests being processed, including those sent in confidential mode; NEAR AI explained that confidential computing protects against the operator and the host, but not against a vulnerability in the software inside the environment. The attacker then gained access to account data for about 100 users of the legacy Private Chat (emails, names, plans, balances) and to 18 agents on a legacy deployment outside the TEE (chat history, settings, logs); the logs contained two NEAR AI Cloud API keys, which have been revoked. No CVE has been assigned to the vulnerability, and the information was passed to the SGLang project.