Linux Kernel CVE-2026-97497

Requires local access to a system running the Linux kernel amdkfd (AMD KFD GPU-compute) driver on AMD GPU hardware, plus the ability to invoke the KFD checkpoint/restore (CRIU) code path that accepts a caller-supplied restore_sdma_id. Exploitation is limited by needing an AMD compute-capable GPU with the KFD driver active, a process using SDMA queues, and use of the CRIU restore interface - not reachable from default remote or non-GPU workloads. Consistent with the local, low-privilege CVSS vector (AV:L/PR:L); no user interaction required. The unchecked value produces an out-of-bounds index whose dominant realistic effect is kernel memory corruption and panic.

This is a local privilege-boundary memory-safety flaw in the Linux kernel's AMD amdkfd (KFD) GPU-compute driver. The NVD CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H = 7.8) frames it as a full local privilege-escalation-class issue with high confidentiality, integrity, and availability impact. Reading the actual root cause tempers this: allocate_sdma_queue accepts a caller-specified sdma queue id (restore_sdma_id, used only by CRIU checkpoint/restore) and fails to bounds-check it against the maximum number of queues, yielding an out-of-bounds array index. The realistic dominant outcome of an unchecked array index is memory corruption leading to a kernel crash/panic (availability), not reliable C:H/I:H exploitation, so the vendor-triple 'C:H/I:H/A:H' is likely inflated for a hardening/bounds-check fix. There is NO EPSS score, NO CISA KEV listing, and NO public POC provided - exploitation status is 'no public exploit identified at time of analysis.' All signals point to LOW real-world urgency: exploitation requires local access on a machine with an AMD GPU exposing KFD compute plus use of the narrow CRIU restore path. Conflict to flag: the input Tag says 'Information Disclosure,' but the described defect is an out-of-bounds array index (memory corruption / DoS), not a disclosure primitive - the tag is mislabeled. CWE is listed N/A; the true weakness is CWE-129 (improper validation of array index) / CWE-787 (OOB write). Prioritize routine kernel patching, not emergency response.

A local unprivileged-to-low-privileged user on a host with an AMD GPU and the amdkfd compute stack invokes the KFD checkpoint/restore (CRIU) interface, supplying a crafted restore_sdma_id larger than the maximum queue count. Because the value is not bounds-checked, the kernel indexes an SDMA queue array out of bounds, corrupting kernel memory and typically causing a panic. No public exploit code is identified at time of analysis; realistic impact is a local denial of service rather than confirmed code execution.