PT-2026-50472

CVE-2026-53923

·

Published

2026-06-17

·

Updated

2026-07-13

Description by

CVSS 3.1
7.5
High
MetricsAV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N
Name of the Vulnerable Software and Affected Versions vLLM versions 0.5.5 through 0.23.1rc0
Description Integer truncation of tensor dimensions in GGUF dequantize kernels within csrc/quantization/gguf/gguf kernel.cu leads to partial tensor processing. The output tensor is allocated at full size using torch::empty, which creates uninitialized memory, but the dequantize CUDA kernel processes only a truncated number of elements because the element count parameter k in the to cuda ggml t function pointer is defined as a 32-bit int. When the product of dimensions m * n exceeds INT MAX, the truncated value is used as the kernel launch grid size, leaving the remaining portion of the tensor unfilled.
In multi-tenant inference deployments, this residual GPU memory may contain tensor data from other users' inference requests, resulting in information disclosure. This issue affects the following functions in csrc/quantization/gguf/gguf kernel.cu:
  • ggml dequantize()
  • ggml mul mat vec a8()
  • ggml mul mat a8()
  • ggml moe a8()
Recommendations Update vLLM to version 0.23.1rc0 or later.
Exploitability
Exploit
Fix
Impacts
Researchers
Vendors
Affected products
Related identifiers
GHSA-5JV2-G5WQ-CMR4
PYSEC-2026-3403
Found an issue in the description? Have something to add? Feel free to write us 👾