CVE-2026-53375: Linux Kernel AMDGPU VCE Vulnerability

CVE-2026-53375 Overview

CVE-2026-53375 is a Linux kernel vulnerability in the AMD GPU Video Coding Engine (VCE) driver (drm/amdgpu/vce). The flaw stems from improper handling of address patching when only one of the low or high address components is valid. In that case, the driver can write a malformed address into the VCE firmware command stream. A local, authenticated attacker with access to the GPU device interface can trigger the condition and influence firmware execution, leading to memory corruption impacting confidentiality, integrity, and availability.

Critical Impact
A local user with GPU access can cause the amdgpu VCE driver to submit a bad address to firmware, resulting in kernel memory corruption and potential privilege escalation on affected Linux systems.

Affected Products

Linux kernel builds including the drm/amdgpu driver with VCE support
Distributions shipping vulnerable stable branches referenced in upstream commits
Systems using AMD GPUs that expose the VCE encode/decode command submission path

Discovery Timeline

2026-07-19 - CVE-2026-53375 published to NVD
2026-07-20 - Last updated in NVD database

Technical Details for CVE-2026-53375

Vulnerability Analysis

The Linux kernel's amdgpu driver patches command buffers submitted to the VCE firmware to translate guest-provided addresses into GPU-visible physical addresses. Address patching operates on 64-bit values expressed as separate low (lo) and high (hi) 32-bit registers. The vulnerable code path did not validate that both halves of the address were present and valid before performing the patch. When only one half was marked valid, the driver combined it with stale or zero data, producing an incorrect address that was then written into firmware command memory. The VCE firmware later dereferences this malformed address, causing kernel-side memory corruption in a security-relevant context. Because the driver runs in kernel mode and the scope is changed per the CVSS vector, exploitation can cross privilege boundaries.

Root Cause

The root cause is missing input validation in the VCE command-stream patcher. The routine that reconstructs 64-bit device addresses accepted partial lo/hi patches, permitting a bad address to be written into firmware structures. This is an Improper Input Validation flaw affecting a kernel driver.

Attack Vector

Exploitation requires local access with permission to open the amdgpu device node and submit VCE command buffers. An attacker crafts a command stream that supplies only one of the low or high address components with the validity marker set. The driver processes the partial patch and writes a corrupted address into the firmware queue, which the VCE microcontroller then uses. No user interaction is required. Technical details are available in the upstream fix commits, including drm/amdgpu/vce fix commit 0ee1715 and additional stable backports.

Detection Methods for CVE-2026-53375

Indicators of Compromise

Kernel log entries from the amdgpu driver referencing VCE ring faults, IB submission errors, or firmware hangs
GPU reset events (amdgpu: GPU reset begin) correlated with unprivileged user processes submitting VCE workloads
Unexpected dmesg traces showing vce_v*_ring or amdgpu_vce_cs_parse warnings

Detection Strategies

Inventory running kernel versions and compare against the fixed commits listed in the upstream references to identify unpatched hosts
Monitor for local processes opening /dev/dri/renderD* and submitting VCE command buffers from non-graphics workloads
Correlate GPU driver crashes with process ancestry to identify potential exploitation attempts against the amdgpu VCE path

Monitoring Recommendations

Forward kernel ring buffer messages to a centralized logging platform and alert on repeated amdgpu VCE parse errors or GPU resets
Track uid/gid of processes accessing DRM render nodes and flag unusual users invoking video encode/decode APIs
Baseline normal VA-API and VCE usage per host so anomalous command submission volume can be detected

How to Mitigate CVE-2026-53375

Immediate Actions Required

Apply the upstream Linux kernel patches referenced by the six stable commits, or install a distribution kernel update that includes them
Restrict access to /dev/dri/renderD* and /dev/dri/card* nodes to trusted users via group membership and udev rules
Reboot affected systems after patch installation to load the fixed kernel

Patch Information

The issue is resolved upstream in the Linux kernel via commits 0ee1715, 2d66d1f, 944db9c, b3d1a0a, de2a02c, and ea2c554. These patches require both lo and hi halves of an address to be valid before the patcher writes the resulting value into the VCE command buffer.

Workarounds

Blacklist or unload the amdgpu VCE functionality on servers that do not require GPU video encode/decode
Remove untrusted users from the video and render groups to prevent submission of VCE command buffers
Use container and namespace policies (seccomp, device cgroups) to deny access to DRM render nodes for workloads that do not need them

bash

# Restrict DRM render node access to a dedicated group
sudo groupadd -f gpu-users
sudo chgrp gpu-users /dev/dri/renderD128
sudo chmod 0660 /dev/dri/renderD128

# Verify running kernel includes the fix
uname -r
grep -m1 'amdgpu_vce_cs_reloc' /proc/kallsyms