CVE-2026-68380 Overview
CVE-2026-68380 is a use-after-free vulnerability in the Linux kernel's accel/amdxdna driver, which provides support for AMD XDNA neural processing units. The flaw exists in the command submission path, where amdxdna_cmd_submit() stores a pointer to current->mm in job->mm without acquiring any reference on the mm_struct. The DRM scheduler worker thread later dereferences job->mm from aie2_sched_job_run(). Because no structural reference is held, the mm_struct can be freed before the scheduler executes the job, resulting in a use-after-free condition [CWE-416].
Critical Impact
A local authenticated user submitting jobs to the AMD XDNA accelerator can trigger a use-after-free of mm_struct, leading to memory corruption, privilege escalation, or kernel crash.
Affected Products
- Linux kernel versions containing the accel/amdxdna driver prior to the fix commits
- Systems using AMD XDNA neural processing unit (NPU) hardware
- Distributions shipping affected mainline and stable kernel branches
Discovery Timeline
- 2026-08-10 - CVE-2026-68380 published to NVD
- 2026-08-13 - Last updated in NVD database
Technical Details for CVE-2026-68380
Vulnerability Analysis
The AMD XDNA accelerator driver (accel/amdxdna) manages job submissions to AMD's on-chip neural processing units through the Direct Rendering Manager (DRM) scheduler framework. When a userspace process submits a command, amdxdna_cmd_submit() records the caller's memory descriptor by copying current->mm into the job structure's mm field. This assignment is a raw pointer copy with no reference counting.
Job execution is asynchronous. The DRM scheduler dispatches queued jobs on a dedicated worker thread that eventually calls aie2_sched_job_run(), which dereferences job->mm to access the submitting process's address space. Between submission and execution, the submitting task can exit and its mm_struct can be released, leaving job->mm dangling.
When the scheduler worker subsequently accesses the freed mm_struct, the kernel operates on memory that may have been reallocated for unrelated data. This produces classic use-after-free consequences, including kernel memory corruption and controllable kernel state that can be leveraged for privilege escalation.
Root Cause
The root cause is missing reference counting on a shared kernel object crossing an asynchronous boundary. The mm_struct structure supports two reference types: mm_users for address-space users and mm_count for structural references. Neither was acquired by amdxdna_cmd_submit(), violating the kernel's ownership contract for cross-thread pointer sharing.
Attack Vector
Exploitation requires local access with permission to open the amdxdna DRM device node and submit commands. An attacker submits a job, then arranges for the submitting task's mm_struct to be released before the scheduler worker runs the job. Winning this race and controlling the freed slab allocation enables write or read primitives against kernel memory. The fix, present in commits 6875ee2bef48, e8fadbffc19a, and faebb7ba1ac6, calls mmgrab() at submission time to acquire an mm_count reference and pairs each cleanup path with mmdrop().
No public proof-of-concept exploit code is available. Refer to the kernel commit for the authoritative patch.
Detection Methods for CVE-2026-68380
Indicators of Compromise
- Kernel Address Sanitizer (KASAN) reports referencing mm_struct, amdxdna_cmd_submit, or aie2_sched_job_run in the call trace
- Unexpected kernel oops or general protection faults originating from the amdxdna DRM scheduler worker
- Slab corruption warnings following heavy job submission activity on AMD XDNA hardware
Detection Strategies
- Enable CONFIG_KASAN on test kernels to surface use-after-free conditions during amdxdna job execution
- Audit dmesg and journal logs for stack traces implicating the DRM scheduler and amdxdna driver
- Monitor kernel version and driver commit hashes across fleet inventory to identify unpatched systems
Monitoring Recommendations
- Collect kernel crash telemetry and correlate crashes against workloads that exercise AMD NPU hardware
- Track processes opening /dev/accel/accel* device nodes and submitting jobs to the amdxdna driver
- Alert on repeated kernel panics or BUG: messages that reference mmdrop, mmput, or DRM scheduler symbols
How to Mitigate CVE-2026-68380
Immediate Actions Required
- Apply the upstream fix by updating to a kernel that includes commits 6875ee2bef48, e8fadbffc19a, or faebb7ba1ac6
- Restrict access to /dev/accel/* device nodes to trusted users through group ownership and permissions
- Prioritize patching on systems equipped with AMD XDNA NPUs where local users can submit compute workloads
Patch Information
The fix acquires a structural reference on the mm_struct at job submission via mmgrab() and releases it with mmdrop() on every cleanup path. Patches are available in the upstream Linux kernel through the following commits: 6875ee2bef48, e8fadbffc19a, and faebb7ba1ac6. Consult your distribution's security tracker for backported stable kernel releases.
Workarounds
- Unload the amdxdna kernel module on systems that do not require AMD NPU functionality using modprobe -r amdxdna
- Blacklist the module via /etc/modprobe.d/ to prevent automatic loading at boot
- Limit device node permissions so only privileged service accounts can submit jobs to the accelerator
# Disable the amdxdna driver where NPU functionality is not required
echo 'blacklist amdxdna' | sudo tee /etc/modprobe.d/disable-amdxdna.conf
sudo modprobe -r amdxdna
# Restrict access to DRM accelerator device nodes
sudo chgrp render /dev/accel/accel0
sudo chmod 0660 /dev/accel/accel0
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
