Critical Vulnerability Found in SGLang LLM Inference Framework

A critical vulnerability, CVE-2026-86793, has been disclosed in the SGLang open-source LLM inference framework. This flaw, a SafeUnpickler bypass, enables attackers to execute arbitrary functions due to an overly broad allowlist and incomplete denylist. No vendor patch is currently available, marking the fourth critical vulnerability in the AI stack within 18 days and highlighting increasing targeting of AI inference infrastructure.