CVE-2026-68106

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 flaw is reached only via the AMDGPU_CS DRM ioctl on a local render node (/dev/dri/renderD*), traversing amdgpu_cs_ioctl, amdgpu_cs_patch_ibs, amdgpu_uvd_ring_parse_cs, and amdgpu_uvd_cs_msg_decode; it is not reachable from any network-facing kernel service.

AC:LAttack complexity, low

An attacker with render-node access fully controls the UVD decode message fields (stream_type H264/H264 Perf, width, height, level) in their own BO and can reliably trigger width or height < 16 after creating a valid UVD session handle.

PR:LPrivileges required, low

Exploitation requires only standard unprivileged local access to the AMDGPU DRM render node (DRM_AUTH|DRM_RENDER_ALLOW on AMDGPU_CS), not root or capabilities outside the render/video group; user namespaces do not grant this access by themselves.

UI:NUser interaction, none

No victim interaction is required; a local attacker can directly submit a crafted UVD command stream through AMDGPU_CS ioctl without needing another user to open media, mount a filesystem, or take any other action.

S:UScope, unchanged

Impact is confined to the kernel/GPU driver security domain on the local host (kernel oops or GPU memory corruption within the attacker's DRM client); it does not cross VM, IOMMU, or sandbox boundaries to affect a different security authority.

C:HConfidentiality impact, high

On x86, kernel-mode divide-by-zero causes a fatal trap; on architectures where unsigned division by zero returns zero, underestimated DPB buffer sizes can pass validation, enabling GPU out-of-bounds reads of adjacent mapped buffers that may hold sensitive data.

I:HIntegrity impact, high

Underestimated buffer-size calculations from zero macroblock counts can let undersized DPB buffers pass kernel validation, permitting the UVD engine to perform out-of-bounds writes into adjacent GPU-mapped memory that could be leveraged for further memory corruption.

A:HAvailability impact, high

Triggering the divide-by-zero in kernel context during CS parsing causes an x86 divide-error exception leading to die()/kernel oops, crashing or hanging the system; repeated submissions can deny GPU/video availability on affected UVD hardware (v3.1-v6.0).