Linux Kernel CVE-2026-68447

Exploitation requires all of the following: (1) AMD v9 architecture GPU hardware present and the amdkfd kernel module loaded (e.g., AMD Vega, CDNA, or Instinct MI300/multi-XCC v9.4.3); (2) local access with at least low privilege (PR:L per CVSS vector), specifically membership in the OS group permitted to open /dev/kfd (typically 'video' or 'render'); (3) ability to create an AMD GPU compute queue and invoke the CRIU checkpoint ioctl path, which is a non-default, GPU-process-migration-specific workflow rather than ordinary GPU rendering or compute use; (4) a hardware-reported cp_hqd_cntl_stack_size that exceeds the allocated BO size, which may be hardware-model-specific. Exploitation is limited to information disclosure from adjacent GTT memory - no write primitive is described. Systems lacking AMD v9 GPU hardware, running amdkfd blacklisted, or where untrusted users are denied /dev/kfd access are not affected.

The NVD CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:N/A:H) scores 7.1 and assigns high availability impact, which is not clearly supported by the description - the described impact is a kernel data leak, not a crash or denial-of-service. The A:H metric warrants verification with the vendor; it may reflect risk of kernel oops from reading unmapped adjacent pages, but this is inferred, not confirmed by the description. The EPSS score of 15% is elevated relative to typical kernel information-disclosure CVEs, likely driven by the Linux kernel's high CVE volume rather than confirmed targeting. No active exploitation (KEV) and no public exploit code have been confirmed. Real-world risk is meaningfully constrained: exploitation requires co-location of AMD v9 GPU hardware, the amdkfd module loaded, local user access to the /dev/kfd device node (typically restricted to the 'video' or 'render' group), and the ability to exercise CRIU GPU queue checkpointing - a relatively niche capability. This is a genuine kernel data-leak requiring patching, but prioritization below actively exploited or remotely triggerable flaws is appropriate.

A local user with membership in the 'video' or 'render' group on a system running AMD v9 GPU hardware creates a compute queue via the AMDKFD ioctl interface and triggers a CRIU checkpoint operation on that queue. The hardware-reported cp_hqd_cntl_stack_size is larger than the queue's allocated control stack BO, causing the kernel's memcpy to read beyond the BO boundary into adjacent GTT memory. The excess bytes - potentially containing kernel stack data, heap structures, or data from adjacent GPU memory allocations - are returned to the attacker's userspace checkpoint buffer without kernel sanitization. No public exploit code has been identified at time of analysis.