CVE-2026-22807
vLLM Dynamic Module Loading Vulnerability Allows Arbitrary Code Execution
Halo Surface Signal: 3 out of 5 — possibly public-facing.
The vLLM inference engine has a vulnerability that allows arbitrary code execution on the host during model loading if it loads code from an attacker-controlled source. This can happen before any requests are processed and does not require API access, posing a risk to systems hosting large language models.