Linux Kernel CVE-2026-68258

Exploitation requires a local user account with at least standard (non-root) privileges on a Linux host where all of the following are true: AMD GPU hardware is present and operational, the amdkfd kernel module is loaded, and the /dev/kfd device node is accessible to the attacking user (typically requires membership in the 'render' or 'video' group, or equivalent DAC permission). The vulnerable code path is exclusively the KFD CRIU restore ioctl - CRIU must be invocable on the system. Systems without AMD GPU hardware, with amdkfd unloaded, or where /dev/kfd access is restricted to root are entirely unexposed. The CVSS PR:L confirms low-privilege local access is sufficient; no authentication beyond local account access is required once device node permissions are met. The absence of network reachability (AV:L) is a hard limiting factor - remote exploitation is not possible.

The CVSS 3.1 vector AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:N/A:H scores 7.1 and accurately captures the local nature of the exploit, the low privilege required to reach the CRIU restore ioctl, and the dual impact of kernel memory disclosure and availability loss (kernel panic). No integrity impact is assessed, consistent with a read-side out-of-bounds flaw. EPSS at 0.17% (7th percentile) is very low and aligns with the absence of any CISA KEV listing or publicly known exploit code, suggesting this has not attracted active exploitation tooling. The narrow attack surface - a Linux host with AMD GPU hardware, the amdkfd module loaded, and CRIU in use - further depresses real-world risk below what the raw CVSS score implies; even a mid-sized municipal IT team in walbrzych running only Intel or NVIDIA GPU workloads faces zero exposure. CVSS confirms the local-only vector, and no conflicting signals suggest network reachability. The primary risk population is HPC clusters, GPU workstations, and data center hosts using AMD GPUs with CRIU-based workload migration. CWE assignment is not confirmed by NVD, so the precise root cause class remains inferred.

A local user on a workstation or HPC node with an AMD GPU invokes a crafted CRIU restore operation targeting KFD GPU queue state, embedding an out-of-range queue_type integer or an unexpected mqd_size in the restore private data. The kernel processes this input without bounds checking, causing a kernel out-of-bounds memory read that either leaks sensitive kernel memory contents back to userspace or triggers a NULL dereference or panic, crashing the system. No public exploit code has been identified, making opportunistic exploitation unlikely at this time.