CVE-2026-45980: Linux Kernel Use-After-Free Vulnerability

CVE-2026-45980 Overview

CVE-2026-45980 is a use-after-free vulnerability in the Linux kernel's accel/amdxdna driver, which supports AMD AI Engine (XDNA) accelerator hardware. The flaw occurs when jobs run on a hardware context while that same context is releasing its resources through aie2_release_resource(). This race between job execution and resource teardown can dereference freed memory and crash the kernel. Upstream maintainers resolved the issue by stopping job scheduling before aie2_release_resource() is invoked and restarting it after release completes. The patched aie2_sched_job_run() now also verifies that the hardware context is still active before executing.

Critical Impact
A race condition between job scheduling and hardware context resource release in the AMD XDNA accelerator driver leads to use-after-free, kernel memory corruption, and potential local denial of service.

Affected Products

Linux kernel builds including the accel/amdxdna driver prior to the fix commits
Systems using AMD XDNA AI Engine accelerators with the in-tree kernel driver
Distributions packaging affected stable kernel branches before backport

Discovery Timeline

2026-05-27 - CVE-2026-45980 published to NVD
2026-05-27 - Last updated in NVD database

Technical Details for CVE-2026-45980

Vulnerability Analysis

The vulnerability lives in the AMD XDNA accelerator driver under accel/amdxdna, which manages hardware contexts (HWCTX) used to dispatch compute jobs to the AI Engine. When a hardware context is torn down, the driver invokes aie2_release_resource() to free associated resources. Prior to the fix, the scheduler could continue submitting jobs to that context while release was in progress.

The scheduler path aie2_sched_job_run() could therefore operate on a context whose backing resources had already been freed. Accessing those freed structures produces a use-after-free condition [CWE-416], which on the kernel side can result in memory corruption, oops, or exploitation primitives depending on heap state.

The upstream fix introduces two coordinated changes. Job scheduling is explicitly stopped before calling aie2_release_resource() and resumed only after the release completes. Additionally, aie2_sched_job_run() now checks whether the hardware context is still active before dispatching work.

Root Cause

The root cause is missing synchronization between the job submission path and the resource release path for an AMD XDNA hardware context. Resource release was not gated against in-flight job runs, allowing the scheduler to touch freed objects.

Attack Vector

Exploitation requires local access to a system with the affected amdxdna driver and the ability to submit jobs to an XDNA hardware context while triggering release of that context. A local user able to interact with the accelerator device can race context teardown against job submission to reach the use-after-free.

No verified public exploit code is available. See the upstream commits 688c3ff079b1, b79d31dce49b, and f1370241fe80 for the canonical fix. References: Kernel Patch 688c3ff079b1, Kernel Patch b79d31dce49b, Kernel Patch f1370241fe80.

Detection Methods for CVE-2026-45980

Indicators of Compromise

Kernel oops or panic traces referencing aie2_sched_job_run, aie2_release_resource, or amdxdna driver symbols
KASAN reports flagging use-after-free in the accel/amdxdna code paths
Unexpected accelerator device resets or hardware context teardown errors in dmesg

Detection Strategies

Inventory running kernels and compare against fixed stable releases that include commits 688c3ff079b1, b79d31dce49b, and f1370241fe80
Enable KASAN on test kernels to surface use-after-free conditions in the XDNA driver before production deployment
Monitor host telemetry for repeated amdxdna related crashes that may indicate exploitation attempts or instability

Monitoring Recommendations

Forward kernel logs (dmesg, journalctl -k) to a centralized log pipeline and alert on amdxdna warnings or oops events
Track which workloads access /dev/accel/* device nodes and correlate with crash events
Watch for local users invoking accelerator ioctls in rapid context create and destroy patterns consistent with race exploitation

How to Mitigate CVE-2026-45980

Immediate Actions Required

Apply the upstream Linux kernel patches referenced in the NVD entry and rebuild or update the kernel package
Restrict access to the AMD XDNA accelerator device nodes to trusted users and processes only
Audit container and VM configurations that pass through the accelerator to ensure untrusted workloads cannot reach the device

Patch Information

The fix is committed to the upstream Linux kernel in three stable-tree commits: 688c3ff079b10e4600f040944430d3d4ff448a15, b79d31dce49b50c79620389b3639280802a86960, and f1370241fe8045702bc9d0812b996791f0500f1b. Distribution maintainers should backport these commits or upgrade to a stable kernel release that contains them. See Kernel Patch 688c3ff079b1, Kernel Patch b79d31dce49b, and Kernel Patch f1370241fe80.

Workarounds

Unload the amdxdna kernel module on systems where the AI Engine accelerator is not required
Use udev rules or device permissions to restrict /dev/accel/* access to a dedicated trusted group
Disable XDNA-related workloads until the kernel is patched on shared multi-tenant systems

bash

# Configuration example: prevent amdxdna from loading until patched
echo 'blacklist amdxdna' | sudo tee /etc/modprobe.d/blacklist-amdxdna.conf
sudo update-initramfs -u
# Verify the module is not loaded after reboot
lsmod | grep amdxdna