CVE-2025-66448

vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.11.1, vllm has a critical remote code execution vector in a config class named Nemotron_Nano_VL_Config. When vllm loads a model config that contains an auto_map entry, the config class resolves that mapping with get_class_from_dynamic_module(...) and immediately instantiates the returned class. This fetches and executes Python from the remote repository referenced in the auto_map string. Crucially, this happens even when the caller explicitly sets trust_remote_code=False in vllm.transformers_utils.config.get_config. In practice, an attacker can publish a benign-looking frontend repo whose config.json points via auto_map to a separate malicious backend repo; loading the frontend will silently run the backend’s code on the victim host. This vulnerability is fixed in 0.11.1.

An adversary registers two GitHub/Hugging Face accounts. The first hosts a legitimate-looking multimodal model repository (frontend repo) with a well-crafted README, model card, and config.json. The config.json includes an auto_map field pointing to the adversary's second repository (backend repo) which hosts a malicious Python class. The frontend repo is promoted in AI/ML communities, referenced in blog posts, or submitted to model leaderboards to build credibility. A target organization's MLOps pipeline or developer runs vllm.LLM('attacker/benign-model', trust_remote_code=False) — the False flag is ignored, vLLM resolves get_class_from_dynamic_module against the auto_map URL, fetches the malicious Python from the backend repo, and executes it on the inference host. The payload can drop a reverse shell, exfiltrate environment variables (AWS credentials, OpenAI keys, internal API tokens), or install a persistent backdoor. The entire compromise happens silently before any inference request is processed.