CVE-2026-41523 is a code injection vulnerability (CWE-94) in vLLM, an inference and serving engine for large language models. Before version 0.22.0, an assert-based security check in the activation function loading can be bypassed when vLLM runs in Python optimized mode (python -O or PYTHONOPTIMIZE=1). This bypass enables any unauthenticated attacker to achieve arbitrary code execution on the server by publishing a malicious HuggingFace model. The vulnerability is addressed in vLLM 0.22.0. The CVSS 3.1 base score is 7.5, indicating high severity with network attack vector, high impact on confidentiality, integrity, and availability, and requiring user interaction with high attack complexity.

An attacker can remotely execute arbitrary code on the server running vulnerable versions of vLLM in Python optimized mode by supplying a malicious model. This compromises the confidentiality, integrity, and availability of the affected system. No known exploits in the wild have been reported yet.

Upgrade vLLM to version 0.22.0 or later, where this vulnerability is fixed. Since this is a code execution vulnerability triggered by running vLLM in Python optimized mode, avoid running vLLM with python -O or PYTHONOPTIMIZE=1 if upgrading is not immediately possible. Patch status is confirmed fixed in 0.22.0 as per the vendor advisory.