CVE-2026-68447: 
Linux Kernel vulnerability analysis and mitigation

CVE-2026-68447 is a memory disclosure vulnerability in the Linux kernel's drm/amdkfd driver affecting the CRIU (Checkpoint/Restore In Userspace) checkpoint mechanism for AMD GPU queues. The flaw causes the kernel to copy the MQD (Memory Queue Descriptor) control stack using the hardware-reported cp_hqd_cntl_stack_size value without bounding it to the allocated buffer object (BO) size, allowing reads past the BO boundary into adjacent GTT memory and leaking kernel data to userspace. It was published on August 12, 2026, and affects Linux kernel versions prior to 7.1.6 (stable branch) and prior to 7.2-rc2 (mainline). The CVSS category is estimated as Medium, with an EPSS score of 0.0 (GitHub Advisory, Feedly).

The root cause is an out-of-bounds read (related to CWE-125) in the drm/amdkfd driver's CRIU checkpoint path. When checkpointing a v9 GPU queue, the code uses cp_hqd_cntl_stack_size — a value read directly from hardware registers — as the length argument to memcpy without validating that it does not exceed the page-aligned size of the allocated control stack BO. If the hardware field is larger than the actual allocation, memcpy reads past the BO boundary into adjacent GTT (Graphics Translation Table) memory, which may contain sensitive kernel data. The fix stores the page-aligned control stack BO size in mqd_manager and clamps all checkpoint copies and reported sizes to min(cp_hqd_cntl_stack_size, mm->ctl_stack_size), with the same bound applied to the multi-XCC v9.4.3 checkpoint layout. The upstream fix is cherry-picked from commit 6c2abd0ec09e86c6323010673766f76050e28aa3 (GitHub Advisory).

A local user with access to the CRIU checkpoint functionality on a system with an AMD GPU can trigger the out-of-bounds read to leak sensitive kernel memory contents to userspace. The primary impact is a confidentiality breach — arbitrary kernel data from adjacent GTT memory regions may be exposed, potentially including cryptographic material, pointers, or other privileged information. Integrity and availability are not directly affected by this vulnerability, and exploitation is limited to local attackers with the necessary privileges to invoke CRIU checkpoint operations on AMD GPU queues (Feedly, GitHub Advisory).

There is no public proof-of-concept exploit and no evidence of in-the-wild exploitation as of the disclosure date. The vulnerability requires local access and the ability to invoke CRIU checkpoint operations on an AMD GPU queue, significantly limiting the attack surface. The EPSS score is 0.0, and the CVE is not listed in the CISA Known Exploited Vulnerabilities (KEV) catalog. No threat actor attribution has been reported (Feedly).

The fix is available in Linux kernel stable version 7.1.6 and mainline 7.2-rc2, via commits a0d87beb2660a5098b2b0ecdc1e96810a9074ea9 and 426ffae6ecc7ec77d32bf8be065c21a1b881b084 respectively. Administrators should update to a patched kernel version as soon as it is available for their distribution. As a short-term workaround, restricting local user access to CRIU checkpoint functionality for AMD GPU queues (e.g., via access controls or disabling CRIU for GPU workloads) can reduce exposure until a patched kernel is deployed (GitHub Advisory, Feedly).