NEAR AI breached through an SGLang RCE — patch your inference servers
An attacker ran code on a live NEAR AI inference box via SGLang, reaching ~100 users' data — and could watch requests even inside the confidential enclave.

Copy markdown
An SGLang bug ran code on a NEAR AI box
On Nov 8, an attacker exploited an SGLang remote-code-execution flaw to run code on a NEAR AI inference server, reaching ~100 users' names and emails (some with plan and balance data), chat logs from 18 legacy non-TEE agents, and two customers' Cloud API keys, now revoked. NEAR disclosed it Nov 9 and says it has patched every server.
Self-hosting on SGLang? Patch now
SGLang has had a string of critical RCEs this year — CVE-2026-5760 (CVSS 9.8) executes arbitrary Python when a server loads a booby-trapped GGUF and renders its chat template. If you serve models with SGLang, update, pin trusted weights, and never load untrusted GGUF or chat templates from the hub.
Confidential mode isn't a force field
The attacker could observe requests running in NEAR's confidential (TEE) mode, because a bug inside the enclave sees the same plaintext the enclave is processing. Trusted execution guards against outside snooping, not against vulnerable code you run within it — threat-model your "private inference" accordingly.