CVE-2026-63855

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 vulnerability is reached only through the local DRM_AMDGPU_CS ioctl on a /dev/dri/renderD* node; there is no network, Bluetooth, or physical-access attack path to the VCN command submission code.

AC:LAttack complexity, low

On affected AMD Instinct MI100 systems, an attacker with render-node access can reliably submit a VCN enc/dec command buffer with a user-fence chunk using standard libdrm/amdgpu APIs, triggering the malformed 64-bit fence write without races or victim-dependent timing.

PR:LPrivileges required, low

Exploitation requires opening the DRM render node and submitting GPU commands, which needs only local user privileges (typically membership in the render/video group) on multi-tenant GPU cloud instances, workstations, or containers with GPU access—not root or capabilities obtainable only in the init namespace.

UI:NUser interaction, none

No victim interaction is required; the attacker directly crafts and submits the malicious command submission through the render node without needing another user to open files, mount filesystems, or perform any action.

S:UScope, unchanged

The misdirected GPU fence write corrupts memory within the attacker's own GPU virtual address space on the assigned device; it does not cross a VM, container, or IOMMU security boundary to affect another security authority.

C:HConfidentiality impact, high

The bug causes GPU-initiated writes to an incorrectly decoded address, constituting out-of-bounds memory corruption that can be leveraged to read adjacent GPU-mapped buffers and disclose sensitive data from the attacker's or co-resident GPU allocations.

I:HIntegrity impact, high

The hardware fence completion write is an attacker-controlled 64-bit store to a mis-decoded GPU virtual address, providing an arbitrary GPU memory write primitive that could corrupt kernel-visible GTT mappings or GPU state for further exploitation.

A:HAvailability impact, high

Misdirected fence writes to unmapped or invalid GPU addresses can cause GPU faults, engine hangs, or full GPU resets that deny video codec and compute service on the affected accelerator.