CVE-2026-53375

Quoted from the kernel CNA's own scoring file, word for word. A highlighted metric carries the most severe value CVSS v3.1 defines for it, which is what pushes the score up. source

AV:LAttack vector, local

The vulnerable code is reached only through the AMDGPU_CS DRM ioctl on a render node (/dev/dri/renderD*), requiring local access to submit crafted VCE indirect buffers to the GPU driver.

AC:LAttack complexity, low

An attacker fully controls the VCE command stream and per-command length fields, and can reliably craft commands where exactly one relocation index is out of bounds (e.g., 0x0500000d with len=32) to trigger partial address patching.

PR:LPrivileges required, low

Exploitation requires only unprivileged local access to the amdgpu DRM render node; AMDGPU_CS is gated by DRM_AUTH|DRM_RENDER_ALLOW, not root or special capabilities.

UI:NUser interaction, none

No victim interaction is required; the attacker drives the entire malicious command submission through the render node.

S:CScope, changed

Partial address patching causes VCE firmware to DMA using attacker-influenced GPU addresses outside the intended buffer object boundaries, crossing the GPU memory/DMA isolation boundary that the relocation logic is meant to enforce.

C:HConfidentiality impact, high

VCE firmware performs DMA reads using the malformed 64-bit address, enabling arbitrary reads from GPU-accessible memory regions (including GTT-mapped system memory) beyond the validated buffer scope.

I:HIntegrity impact, high

VCE firmware performs DMA writes (encoded bitstreams, feedback, and context buffers) to attacker-influenced addresses, enabling arbitrary GPU-mapped memory corruption and potential privilege escalation primitives.

A:HAvailability impact, high

Submitting commands with malformed firmware addresses can cause GPU page faults, driver hangs, GPU resets, or kernel oops/panic when the VCE engine accesses invalid memory regions.