PT-2025-47648 · Pytorch+1 · Pytorch+1
Published
2025-11-20
·
Updated
2026-01-10
·
CVE-2025-62164
CVSS v2.0
9.0
High
| Vector | AV:N/AC:L/Au:S/C:C/I:C/A:C |
Name of the Vulnerable Software and Affected Versions
vLLM versions 0.10.2 through 0.11.0
Description
vLLM is an inference and serving engine for large language models (LLMs). A memory corruption issue exists in the Completions API endpoint, specifically when processing user-supplied prompt embeddings. The endpoint utilizes
torch.load() without adequate validation, and due to changes in PyTorch 2.8.0 disabling sparse tensor integrity checks by default, maliciously crafted tensors can bypass security measures. This can lead to out-of-bounds memory writes during the to dense() function call, potentially causing a crash (denial-of-service) and remote code execution (RCE) on the server. The vulnerability is triggered by loading serialized tensors without sufficient validation.Recommendations
Update vLLM to version 0.11.1.
Exploit
Fix
DoS
RCE
Deserialization of Untrusted Data
Memory Corruption
Found an issue in the description? Have something to add? Feel free to write us 👾
Related Identifiers
Affected Products
Pytorch
Vllm