Nixpkgs security tracker

Untriaged

Permalink CVE-2026-53923

5.3 MEDIUM

CVSS version (CVSS): 4.0
Attack Vector (AV): Network (N)
Attack Complexity (AC): Low (L)
Attack Requirement (AT): None (N)
Privileges Required (PR): None (N)
User Interaction (UI): Passive (P)
Vulnerable System Impact Confidentiality (VC): Low (L)
Vulnerable System Impact Integrity (VI): Low (L)
Vulnerable System Impact Availability (VA): None (N)
Subsequent System Impact Confidentiality (SC): None (N)
Subsequent System Impact Integrity (SI): None (N)
Subsequent System Impact Availability (SA): None (N)
Modified Attack Vector (MAV): Network (N)
Modified Attack Complexity (MAC): Low (L)
Modified Attack Requirement (MAT): None (N)
Modified Privileges Required (MPR): None (N)
Modified User Interaction (MUI): Passive (P)
Modified Vulnerable System Impact Confidentiality (MVC): Low (L)
Modified Vulnerable System Impact Integrity (MVI): Low (L)
Modified Vulnerable System Impact Availability (MVA): None (N)
Modified Subsequent System Impact Confidentiality (MSC): Negligible (N)
Modified Subsequent System Impact Integrity (MSI): Negligible (N)
Modified Subsequent System Impact Availability (MSA): Negligible (N)
Safety (S): Not Defined (X)
Automatable (AU): Not Defined (X)
Recovery (R): Not Defined (X)
Value Density (V): Not Defined (X)
Vulnerability Response Effort (RE): Not Defined (X)
Provider Urgency (U): Not Defined (X)
Confidentiality Req. (CR): Not Defined (X)
Integrity Req. (IR): Not Defined (X)
Availability Req. (AR): Not Defined (X)
Exploit Maturity (E): Not Defined (X)

created 3 weeks, 2 days ago Activity log

Created suggestion 3 weeks, 2 days ago

vLLM GGUF Kernels: int64_t to int truncation of tensor dimensions causes GPU buffer overflow

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.