NEAR AI has resumed operations of its cloud inference service after a security breach that compromised its internal infrastructure and bypassed secure enclaves meant to isolate private AI conversations. The company said an attacker exploited a remote code execution flaw in SGLang, an open source inference engine, on Nov 8. While the server was hosted in a confidential computing environment, the firm acknowledged that this technology does not protect against vulnerabilities located within the software running inside the enclave.
The breach resulted in the theft of account details from about 100 users and the chat histories of 18 AI agents, both of which were stored outside the secure enclave. NEAR AI also stated that the intruder may have seen confidential requests processed on the affected server during the attack. The firm has since rotated affected credentials and patched the SGLang bug across its servers, noting that the NEAR Protocol was not impacted by the incident.
