Linux Kernel EUVD-2026-59754

Exploitation requires all of the following: (1) a local user account on the target Linux system; (2) membership in the group with read-write access to /dev/kfd (typically 'render' or 'video'), which is granted by default to GPU compute users on ROCm-enabled distributions but is not present on general-purpose servers without AMD GPU compute software installed; (3) an AMD GPU from the GFX8 generation (Fiji: Radeon R9 Fury/Nano, Polaris: RX 400/500 series) or GFX9 generation (Vega: RX Vega, Radeon VII, Instinct MI series) physically installed and driven by amdkfd; (4) an affected kernel version as listed above. Remote exploitation is not possible under any configuration - the attack surface is entirely local via kernel ioctl. The limiting factors (local access + KFD group membership + specific AMD GPU generation) substantially constrain the realistic attack population to GPU compute workstations and ROCm/ML inference servers.

The CVSS 7.8 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H) correctly characterizes the attack as local with low-privilege entry. Real-world risk is materially lower than the raw score suggests for three reasons. First, exploitation requires specific AMD GFX8/GFX9 GPU hardware installed and the amdkfd driver loaded - this excludes the majority of Linux deployments, which have no AMD compute GPU. Second, the attacker must hold an account with KFD device node access (typically via 'render' or 'video' group membership), a non-default privilege in general-purpose server environments, though common in ROCm/ML workstation setups. Third, EPSS at 0.17% (6th percentile) and absence from CISA KEV both indicate no observed exploitation. The high C/I/A impact ratings are plausible in principle - overrunning kernel memory adjacent to the BO could enable kernel code execution or information disclosure - but converting the overflow to a reliable privilege escalation primitive requires knowledge of adjacent kernel heap layout, adding practical complexity not captured by AC:L. No conflicting signals between CVSS vector and description were detected; the local-only vector is consistent with requiring KFD ioctl submission.

A local user on an AMD ROCm-capable ML workstation or compute server - for example, a member of the 'render' group - submits a GPU queue creation ioctl with eop_ring_buffer_size set to 0xFFFFF001. The 32-bit ALIGN overflow wraps the computed size to zero, causing kfd_queue_buffer_get() to accept a 4 KB buffer object in place of the required 8 KB EOP ring on the system's Vega GFX9 GPU. Subsequent Command Processor EOP writes overrun the allocation into adjacent kernel heap memory, either crashing the GPU driver or providing a foothold for kernel memory corruption leading to privilege escalation.