vllm: Malicious model to RCE by torch.load in hf_model_weights_iterator

CAO 8.8CVE-2025-24357 · GHSA-rh4j-5rhw-hr54 · PYSEC-2025-58 · Công bố 27/01/2025 · cập nhật 17/06/2026

Tóm tắt nhanh

CVE-2025-24357 là lỗ hổng mức CAO (CVSS 8.8) ảnh hưởng vllm, vllm (vllm).
Xác suất bị khai thác trong 30 ngày tới (EPSS): 0.87%, cao hơn 58.8% các lỗ hổng khác.
Cách xử lý: cập nhật lên 0.7.0 hoặc mới hơn.

Gói / sản phẩm	Phiên bản bị ảnh hưởng	Phiên bản đã vá
vllm PyPI	`< 0.7.0`	`0.7.0`
vllm vllm	`< 0.7.0`	`0.7.0`

Mô tả

Description

The vllm/model_executor/weight_utils.py implements hf_model_weights_iterator to load the model checkpoint, which is downloaded from huggingface. It use torch.load function and weights_only parameter is default value False. There is a security warning on https://pytorch.org/docs/stable/generated/torch.load.html, when torch.load load a malicious pickle data it will execute arbitrary code during unpickling.

Impact

This vulnerability can be exploited to execute arbitrary codes and OS commands in the victim machine who fetch the pretrained repo remotely.

Note that most models now use the safetensors format, which is not vulnerable to this issue.

References

Mô tả của NVD

vLLM is a library for LLM inference and serving. vllm/model_executor/weight_utils.py implements hf_model_weights_iterator to load the model checkpoint, which is downloaded from huggingface. It uses the torch.load function and the weights_only parameter defaults to False. When torch.load loads malicious pickle data, it will execute arbitrary code during unpickling. This vulnerability is fixed in v0.7.0.