CVE-2026-22807

vLLM is an inference and serving engine for large language models (LLMs). Starting in version 0.10.1 and prior to version 0.14.0, vLLM loads Hugging Face `auto_map` dynamic modules during model resolution without gating on `trust_remote_code`, allowing attacker-controlled Python code in a model repo/path to execute at server startup. An attacker who can influence the model repo/path (local directory or remote Hugging Face repo) can achieve arbitrary code execution on the vLLM host during model load. This happens before any request handling and does not require API access. Version 0.14.0 fixes the issue.

An adversary creates a malicious model repository on HuggingFace and embeds a reverse shell or credential harvester in the auto_map field of the model's config.json as attacker-controlled Python code. The adversary engineers model adoption via social engineering targeting ML engineers ('try this fine-tuned model'), compromising an upstream model dependency, or gaining write access to a CI/CD pipeline configuration specifying model paths. When the unpatched vLLM instance initializes, it resolves and executes the auto_map module with vLLM process privileges — before serving any requests and before any API-layer security controls engage — achieving full host compromise. No vLLM API credentials are required, only influence over the model path.