vLLM Remote Code Execution Vulnerability via Unchecked Dynamic Module Loading

A remote code execution vulnerability exists in vLLM, an inference and serving engine for large language models. This issue affects vLLM versions 0.10.1 prior to 0.14.0. The vulnerability arises because vLLM loads dynamic modules from Hugging Face's 'auto_map' during model resolution without verifying the 'trust_remote_code' parameter. As a result, attacker-controlled Python code in a model repository or path can be executed on the vLLM host during model loading, before any requests are processed, and without requiring API access.

Impact

Exploitation of this vulnerability allows for arbitrary code execution on the vLLM host during the model loading process.

Remediation

Users can upgrade to vLLM version 0.14.0 or later, where this vulnerability has been fixed.