CVE-2025-23275: Nvidia Cuda Toolkit DOS Vulnerability

CVE-2025-23275 Overview

CVE-2025-23275 affects the NVIDIA CUDA Toolkit across all supported platforms. The vulnerability resides in the nvJPEG library, which handles JPEG image decoding on NVIDIA GPUs. A local authenticated user can trigger a GPU out-of-bounds write by supplying crafted image dimensions to the library. Successful exploitation may lead to denial of service and information disclosure. The flaw is tracked under CWE-787 (Out-of-bounds Write) and impacts CUDA Toolkit deployments on Linux, Windows, DriveOS, and Linux for Tegra.

Critical Impact
Local authenticated attackers can trigger GPU memory corruption in the nvJPEG library, potentially leading to denial of service and disclosure of GPU-resident data.

Affected Products

NVIDIA CUDA Toolkit (all platforms)
NVIDIA DriveOS and NVIDIA Linux for Tegra
NVIDIA nvJPEG library on Linux and Microsoft Windows hosts

Discovery Timeline

2025-09-24 - CVE-2025-23275 published to NVD
2026-06-17 - Last updated in NVD database

Technical Details for CVE-2025-23275

Vulnerability Analysis

The nvJPEG library provides GPU-accelerated JPEG decoding as part of the NVIDIA CUDA Toolkit. The vulnerability occurs when the library processes image dimensions that fall outside expected boundary conditions. When a local authenticated user submits an image with certain crafted dimensions, nvJPEG writes beyond the intended GPU memory allocation. This out-of-bounds write corrupts adjacent memory regions on the device. Because GPU memory can hold sensitive data from concurrent contexts, corruption can cascade into denial of service and leak residual information from GPU buffers. The attack is constrained to local execution but does not require elevated privileges beyond a standard authenticated session.

Root Cause

The root cause is insufficient validation of image dimension parameters prior to GPU buffer indexing inside nvJPEG. The library trusts caller-provided dimensions when computing write offsets in device memory. Missing bounds checks allow indices to exceed allocated buffer sizes, matching the classic CWE-787 pattern.

Attack Vector

Exploitation requires local access with authenticated user privileges. The attacker submits a malformed JPEG or invokes the nvJPEG API with attacker-controlled dimension arguments. No user interaction is needed, and the scope remains unchanged. The result is corruption of GPU memory that can crash CUDA workloads or expose data from other GPU allocations. Refer to the NVIDIA Security Bulletin for technical details specific to the affected components.

Detection Methods for CVE-2025-23275

Indicators of Compromise

Unexpected CUDA process crashes or CUDA_ERROR_ILLEGAL_ADDRESS faults in application logs when handling JPEG inputs.
GPU driver logs or dmesg entries reporting XID errors correlated with nvJPEG-linked workloads.
Repeated submission of JPEG files with anomalous width or height metadata by non-administrative users.

Detection Strategies

Monitor CUDA runtime error codes emitted by applications that link nvJPEG, and alert on repeated illegal memory access faults.
Inspect JPEG inputs at ingestion boundaries for dimension fields that exceed processing policy limits before they reach GPU workloads.
Correlate local user session activity with nvJPEG API usage to identify unusual invocation patterns from non-service accounts.

Monitoring Recommendations

Enable verbose logging in imaging pipelines that use CUDA Toolkit and forward logs to a centralized SIEM for baseline analysis.
Track NVIDIA driver XID events and GPU device resets as operational health signals tied to potential exploitation attempts.
Maintain an inventory of hosts running vulnerable CUDA Toolkit versions to prioritize patch verification and monitoring coverage.

How to Mitigate CVE-2025-23275

Immediate Actions Required

Apply the patched CUDA Toolkit release from NVIDIA as documented in the NVIDIA Security Bulletin.
Restrict local access to systems running CUDA workloads to trusted, authenticated users only.
Audit applications that invoke nvJPEG and enforce validation of image dimensions at the application layer.

Patch Information

NVIDIA has released updated CUDA Toolkit versions for all affected platforms, including Linux, Windows, DriveOS, and Linux for Tegra. Consult the NVIDIA Support Advisory for the specific fixed versions applicable to each product line and deployment target.

Workarounds

Limit execution of untrusted JPEG inputs on shared GPU hosts until patches are deployed across the environment.
Constrain user permissions on CUDA-enabled systems to reduce the population of accounts able to invoke nvJPEG.
Isolate multi-tenant GPU workloads using separate processes or MIG partitions to reduce blast radius of memory corruption.

bash

# Verify installed CUDA Toolkit version on Linux
nvcc --version
cat /usr/local/cuda/version.json | grep -A1 cuda_nvjpeg

# Confirm nvJPEG library version linked in the environment
ldconfig -p | grep libnvjpeg