Linux Kernel EUVD-2026-59753

Exploitation requires all of the following specific conditions: (1) an AMD GPU present on the target system with the drm/amdkfd kernel module loaded and functional; (2) a low-privilege local user account with read-write access to /dev/kfd - typically granted via membership in the 'render' group or equivalent udev-managed group, which is non-default for arbitrary local users but standard for GPU compute users; (3) the ability to initiate a CRIU checkpoint of a process that owns KFD events; (4) concurrent ability to call kfd_event_create() or kfd_event_destroy() from a second thread within the same process during the checkpoint window. CRIU-based GPU process checkpointing is a specialized operational workflow not present in typical Linux deployments - the vast majority of AMD GPU systems do not run this code path at all. Remote exploitation is not possible; the attack vector is strictly local (AV:L). The race window between count and IDR walk may require multiple attempts to win consistently, though no experimental data on window width or reliability is available from public sources.

The NVD-assigned CVSS vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H, score 7.8) rates this as High severity with low attack complexity, but the AC:L rating is independently disputed here: exploiting a kernel TOCTOU race requires precise thread scheduling to win the count-to-walk window, which aligns more accurately with AC:H in standard race-condition exploit practice - operators should verify this assessment against any supplemental vendor guidance. EPSS at 0.17% (6th percentile) signals very low current exploitation probability, with no evidence of organized adversary activity. No CISA KEV entry and no public proof-of-concept code have been identified. Real-world risk is bounded by the requirement for an AMD GPU to be present with the amdkfd module loaded, a process actively undergoing CRIU checkpointing, and a concurrent thread racing KFD event lifecycle operations - a narrow operational window absent from most general-purpose Linux deployments. That said, even a mid-sized enterprise IT team in malbork running AMD GPU clusters for HPC or machine learning pipelines with CRIU-based container live migration should treat the patch cycle as urgent, because the resulting heap corruption primitives - out-of-bounds write and use-after-free - are well-understood kernel exploit building blocks that skilled local attackers have repeatedly weaponized for privilege escalation to root. The AC:L versus AC:H discrepancy is the key unresolved signal in this analysis.

A local attacker with a low-privilege account on a Linux system with an AMD GPU initializes a KFD process that owns multiple events, then spawns a second thread that rapidly and continuously creates and destroys KFD events. Simultaneously, the attacker initiates a CRIU checkpoint of the process, and by winning the race between kfd_get_num_events() and the subsequent IDR fill walk, triggers either a heap out-of-bounds write (if an event is inserted in the window) or a use-after-free (if an event is removed), providing kernel memory corruption primitives for a subsequent local privilege escalation to root. No public proof-of-concept code has been identified at time of analysis.