vLLM affected by RCE via auto_map dynamic module loading during model initialization

vLLM automatically loads and executes Python code from model files without verifying if they are trustworthy, allowing an attacker to run malicious code on the server when a model is loaded. This is dangerous because it happens automatically at startup, before the system is even ready to handle requests.

vLLM versions 0.10.1 through 0.13.x load Hugging Face auto_map dynamic modules during model initialization without enforcing trust_remote_code validation (CWE-94: Improper Control of Generation of Code). An attacker with control over the model repository path (local or remote) can inject arbitrary Python code that executes with server privileges at startup, before request handling begins, requiring no API access.