CVE-2025-62164

[CVE catalog]

Severity

High · CVSS 8.8CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H

Published

Nov 21, 2025, 1:18:38 AM UTC

Updated

Oct 17, 2026, 10:52:20 AM UTC

Source

[NVD] · [CVE record]

EPSS

0.96% · Oct 25, 2026

A vulnerability in vLLM allows attackers to supply malicious serialized prompt-embedding tensors that are deserialized using torch.load() without validation. Due to PyTorch 2.8.0 disabling sparse-tensor integrity checks by default, a crafted tensor can bypass bounds checks and cause an out-of-bounds write during to_dense(), leading to a crash (DoS) and potentially remote code execution on the vLLM server. vllm: VLLM deserialization vulnerability leading to DoS and potential RCE This vulnerability is considered important rather than moderate because it involves unsafe deserialization leading to memory corruption in a network-reachable, unauthenticated API path. Unlike typical moderate flaws that may only allow limited DoS or require specific conditions, this issue allows an attacker to supply a crafted sparse tensor that triggers an out-of-bounds memory write during PyTorch’s to_dense() conversion. Memory corruption in a server process handling untrusted input significantly elevates security risk because it can lead not only to a reliable crash but also to potential remote code execution, enabling full compromise of the vLLM service. Additionally, the affected code path is part of the standard Completions API workflow, making the attack surface broadly exposed in real deployments. The combination of remote exploitability, unauthenticated access, memory corruption, and potential RCE clearly positions this issue above a moderate classification and into an important severity level. The CVSS score(s) listed for this vulnerability do not reflect the associated product's status, and are included for informational purposes to better understand the severity of this vulnerability.