Remote Code Execution Vulnerability in vLLM by vllm Project
CVE-2025-66448
7.1HIGH
What is CVE-2025-66448?
vLLM, an inference and serving engine for large language models, contains a significant vulnerability in versions prior to 0.11.1. This issue arises from the configuration class Nemotron_Nano_VL_Config which exposes a remote code execution risk when loading a model configuration that features an auto_map entry. Despite the trust_remote_code setting being set to False, the vulnerability allows for the automatic fetching and execution of malicious Python code hosted on a remote repository, paving the way for potential attacks. Users are advised to upgrade to version 0.11.1 to mitigate this risk effectively.
Affected Version(s)
vllm < 0.11.1