CVE-2026-68107

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 only reachable through the local AMDGPU_CS DRM ioctl on /dev/dri/renderD* (amdgpu_cs_ioctl → amdgpu_cs_patch_jobs → amdgpu_cs_patch_ibs), which kmaps the user IB and calls vcn_v4_0_ring_patch_cs_in_place(); no network or physical-device path exists.

AC:LAttack complexity, low

patch_cs_in_place reads IB contents directly from attacker-mapped GEM memory without copying; a concurrent thread can change the param-length dword between vcn_v4_0_enc_find_ib_param()'s initial read and the later idx+=amdgpu_ib_get_value(ib,idx)/4, so the attacker controls both sides of the TOCTOU race.

PR:LPrivileges required, low

AMDGPU_CS is registered with DRM_AUTH|DRM_RENDER_ALLOW, so any local user who can open the AMDGPU render node (typical render/video group on desktops, gaming handhelds, and GPU cloud instances) can submit VCN4 command buffers without root or init-namespace capabilities.

UI:NUser interaction, none

Exploitation requires only the attacker submitting crafted VCN encode IBs through their own DRM context; no separate victim action such as opening a file, mounting a filesystem, or interacting with malicious content is needed.

S:CScope, changed

On VFIO GPU-passthrough and cloud ML hosts, a malicious VM guest submits VCN command streams through the assigned AMD GPU to corrupt host amdgpu kernel state, crossing the guest-to-hypervisor security boundary analogous to a device-model escape.

C:HConfidentiality impact, high

Racing the length dword smaller under-advances IB iteration so nested RADEON_VCN_ENGINE_INFO entries inside a parent package are reached; vcn_v4_0_dec_msg() then reads user-mapped message buffers without snapshotting, enabling TOCTOU check/use mismatches on length/offset fields and out-of-bounds kernel reads past validated bounds.

I:HIntegrity impact, high

The same nested-entry bypass plus unsnapshotted dec_msg parsing lets userspace rewrite offset/size after kernel validation, driving out-of-bounds accesses in the message loop; skipping vcn_v4_0_limit_sched() via mis-iteration also leaves scheduler/BO state inconsistently modified in attacker-influenceable ways.

A:HAvailability impact, high

Racing the length field to zero on re-read makes idx+=0 and re-enters the same ENGINE_INFO entry indefinitely, soft-locking the CS ioctl path; vcn_v4_0_limit_sched() may also block in dma_fence_wait(), yielding repeatable kernel denial-of-service on shared GPU hosts.