On July 19, 2026, the National Vulnerability Database published CVE-2026-53374, a defect in the Linux kernel’s AMDGPU driver that can cause AMD graphics hardware to read stale or uninitialized memory from the GPU’s page table. Kernel maintainers have already issued a fix, and users running Linux on any AMD GPU should move to a patched kernel version—6.1.175, 6.6.140, 6.12.90, 6.18.32, 7.0.9, or newer—without delay.

The Flaw in Detail: Stale GPU Page-Table Entries

The root of the problem lies in how the AMDGPU driver manages the Graphics Address Remapping Table (GART), a hardware structure that maps virtual addresses for GPU workloads. When the driver allocates a new GART table, it creates a buffer that isn’t automatically zeroed out. The translation lookaside buffer (TLB) on the GPU is flushed after unmapping memory, but not after mapping it. As a result, when the driver writes a single page-table entry (PTE), the TLB may speculatively load neighboring, uninitialized entries from the same cache line.

Those garbage entries can appear valid to the GPU, effectively giving it a stale, potentially misleading view of memory. If the driver later writes another PTE in that same cache line, the GPU might continue using the stale entry from its TLB rather than the correct one. In practical terms, this is a memory translation correctness bug—the GPU can be fooled into using an address it shouldn’t, leading to unpredictable behavior, hangs, or data corruption.

What Systems Are Affected?

The vulnerable code lives in drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c, so any Linux system with an AMD GPU that loads the amdgpu driver is potentially affected. That includes:

  • Desktop and workstation users running AMD Radeon graphics on any Linux distribution.
  • GPU compute nodes in data centers, render farms, or machine learning clusters using ROCm or other AMD GPU-accelerated workloads.
  • Virtualized environments that pass an AMD GPU through to a Linux guest (e.g., via Hyper-V DDA, KVM, or Xen).
  • Appliance-style deployments built on embedded or custom Linux kernels that include the amdgpu driver.

Windows systems are not impacted by this CVE. The issue is specific to the open-source AMDGPU kernel driver; Windows uses AMD’s separate Windows kernel-mode display driver. WSL users with standard GPU acceleration are also outside the scope of this advisory, as Microsoft’s WSL GPU virtualization does not expose the vulnerable code path. Nevertheless, IT teams managing mixed Windows-Linux estates should still audit any Linux VMs, dual-boot configurations, or build servers that rely on AMD hardware.

Distribution kernels often backport fixes, so a system’s reported kernel version alone may not tell the whole story. For example, SUSE has already integrated the fix into a June kernel security update, even though its version string may differ from the upstream numbers. Administrators should consult their vendor’s changelog or security notice for precise confirmation.

How the Kernel Fix Addresses the Issue

The patch is elegantly simple: after allocating the GART table, the driver now immediately calls memset_io() to zero-initialize the entire table with the GPU’s standard invalid-or-safe PTE flags (gart_pte_flags). This ensures every entry starts in a known, harmless state before any real mapping is installed.

Why not use AMDGPU’s built-in mechanism for clearing video memory? The AMDGPU_GEM_CREATE_VRAM_CLEARED flag triggers an SDMA engine operation, but the SDMA engine itself depends on a functioning GART table. Initializing the table via SDMA would create a circular dependency. The CPU-side clear, while adding a one-time cost of about 1 MB on allocation, eliminates the bootstrapping headache and guarantees a clean slate.

The Road to This Patch: Timeline and Context

AMD engineer Philip Yang first posted the fix to the AMD graphics mailing list on April 28, 2026, where it was reviewed and merged into the DRM tree. The commit, cherry-picked from d9af8263b82b6eaa60c5718e0c6631c5037e4b24, was then backported across multiple stable kernel series. NVD published the CVE record on July 19, 2026, based on information from kernel.org.

Notably, NVD has not yet assigned a CVSS score, attack vector, or weakness enumeration. There is no evidence of active exploitation in the wild. The CVE is best viewed as a driver correctness defect that has been proactively identified and resolved, rather than a confirmed security incident. Still, its potential to corrupt GPU address translation makes it a priority for any environment where stability and data integrity matter.

What Should You Do Today?

  1. Identify susceptible systems. On any Linux host with an AMD GPU, run lspci | grep -i amd or lsmod | grep amdgpu to confirm the driver is active. Check the running kernel with uname -r.
  2. Update the kernel. If you manage your own kernel, upgrade to one of the fixed upstream versions: 6.1.175+, 6.6.140+, 6.12.90+, 6.18.32+, or 7.0.9+. Most users should rely on their distribution’s package manager. For example, on Debian/Ubuntu: sudo apt update && sudo apt upgrade; on RHEL/Fedora: sudo dnf upgrade kernel.
  3. Reboot into the new kernel. A kernel update is not active until the system restarts. Confirm with uname -r after the reboot.
  4. Verify vendor advisories. If you use a distribution that ships its own kernel (Red Hat, SUSE, Oracle, Ubuntu, etc.), check the security advisory that corresponds to your OS version. The CVE entry provides a list of affected git commits, but your distribution’s changelog is the definitive source for the backported fix.
  5. Don’t panic. There is no indication of active exploitation, and the remediation is a standard kernel update. For Windows-only fleets, no action is required.

Looking Ahead: What to Watch For

NVD may enrich the CVE with a CVSS severity score in the coming weeks, though the fix is already available. The more important signal will come from individual Linux vendors as they publish their own advisories and, potentially, automated detection logic (e.g., OVAL checks). Keep an eye on your distribution’s security mailing list and apply kernel updates through normal patch cycles. For organizations running large AMD GPU clusters, consider adding a check for this CVE to your vulnerability management workflow, but treat it as a driver bug that’s already been put to rest.