CVE-2026-53923
High · 7.5CVSS
7.5
EPSS
0.5%
Published
Jun 22, 2026
CISA due date
—
The record
vLLM is an inference and serving engine for large language models (LLMs). From 0.5.5 until 0.23.1rc0, integer truncation of tensor dimensions in vLLM's GGUF dequantize kernels (csrc/quantization/gguf/gguf_kernel.cu) causes partial tensor processing. The output tensor is allocated at full size via torch::empty (uninitialized memory), but the dequantize CUDA kernel processes only a truncated number of elements. The unfilled portion of the output tensor retains whatever was previously in GPU memory. In multi-tenant inference deployments, this residual GPU memory may contain tensor data from other users' inference requests, constituting information disclosure. This vulnerability is fixed in 0.23.1rc0.
Affected packages
| Ecosystem | Package | Introduced | Fixed |
|---|
Related vulnerabilities
- CVE-2026-114657 Grav from 2.0.0 before 2.2.5 contains an information disclosure vulnerability that allows… CWE-200
- CVE-2026-114574 A security vulnerability has been detected in massCodeIO massCode 5.10. CWE-200
- CVE-2026-111756 follow-redirects is a Node.js HTTP and HTTPS client wrapper that automatically follows… CWE-200
- CVE-2026-111755 follow-redirects is a Node.js HTTP and HTTPS client wrapper that automatically follows… CWE-200
- CVE-2026-112091 A flaw has been found in Dav-ee dairy-management-system… CWE-200
- CVE-2026-109259 ImageMagick before 6.9.13-57 and 7.x before 7.1.2-32 contains an uninitialized memory… CWE-200
- CVE-2026-107383 MariaDB Connector/Node.js is used to connect applications developed on Node.js to MariaDB… CWE-200
- CVE-2026-106120 LiquidJS is a Shopify / GitHub Pages compatible template engine in pure JavaScript. CWE-200