CVE-2026-74716: Linux Kernel Privilege Escalation Vulnerability

CVE-2026-74716 Overview

CVE-2026-74716 is a locally exploitable kernel assertion failure in the Linux kernel accel/amdxdna driver. The flaw resides in amdxdna_insert_pages(), which incorrectly sets VM_MIXEDMAP and clears VM_PFNMAP on GEM object VMAs. An unprivileged local user can trigger a predictable BUG_ON in vmf_insert_pfn_prot() by mmap'ing a non-imported GEM object, calling madvise(MADV_DONTNEED), and then re-accessing the memory. The resulting kernel assertion produces a denial-of-service condition on affected systems running the AMD XDNA accelerator driver.

Critical Impact
Any unprivileged local user can crash the kernel on systems using the AMD XDNA accelerator driver, resulting in a denial-of-service condition.

Affected Products

Linux kernel versions containing the accel/amdxdna driver prior to the fix commits
Systems with AMD XDNA (AI Engine / NPU) accelerator hardware and driver enabled
Distributions shipping the vulnerable amdxdna DRM/accel module

Discovery Timeline

2026-08-22 - CVE-2026-74716 published to NVD
2026-08-22 - Last updated in NVD database

Technical Details for CVE-2026-74716

Vulnerability Analysis

The vulnerability is a locally triggerable kernel BUG_ON [Kernel Vulnerability / Denial of Service] in the AMD XDNA accelerator driver. In amdxdna_insert_pages(), the driver calls vm_flags_mod() to set VM_MIXEDMAP and clear VM_PFNMAP on the target VMA. This flag combination is inconsistent with how the shmem backing pages are subsequently faulted in.

When userspace invokes madvise(MADV_DONTNEED) on the mapping, the kernel clears the page table entries. VM_MIXEDMAP permits this operation, whereas VM_PFNMAP would not. On the next access, drm_gem_shmem_fault() handles the fault and calls vmf_insert_pfn(), which invokes vmf_insert_pfn_prot(). That helper contains the explicit assertion BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)). Because the shmem backing page is normal system memory with a valid pfn, and the VMA now carries VM_MIXEDMAP, the assertion fires and crashes the kernel.

Root Cause

The root cause is an incorrect VMA flag configuration in amdxdna_insert_pages(). The driver pre-populates PTEs with vm_insert_pages() while setting VM_MIXEDMAP, but the fault-time path relies on vmf_insert_pfn(), which requires VM_PFNMAP when the underlying pfn is valid system memory. The mismatch violates a kernel invariant enforced by BUG_ON.

Attack Vector

An attacker with local unprivileged access opens the amdxdna device, mmaps a non-imported GEM object, issues madvise(MADV_DONTNEED) to clear the PTEs, then re-touches the memory. The subsequent page fault path deterministically triggers the BUG_ON, panicking the kernel. No special capabilities or credentials are required beyond access to the amdxdna device node.

No public exploit code is required beyond the sequence described; refer to the upstream fixes at Kernel commit 4a19f7ab and Kernel commit 13339132 for the exact code paths.

Detection Methods for CVE-2026-74716

Indicators of Compromise

Kernel panic or oops messages referencing vmf_insert_pfn_prot, drm_gem_shmem_fault, or amdxdna_insert_pages in dmesg or /var/log/kern.log.
Unexpected system reboots or hangs on hosts equipped with AMD XDNA accelerators.
Unprivileged processes repeatedly opening /dev/accel/accel* device nodes and calling mmap() followed by madvise(MADV_DONTNEED).

Detection Strategies

Monitor kernel ring buffer for BUG: and Call Trace: entries containing the amdxdna or drm_gem_shmem symbols.
Audit process syscall telemetry for the mmap → madvise(MADV_DONTNEED) → memory access sequence against /dev/accel/* handles.
Correlate crash-dump signatures with the assertion string VM_MIXEDMAP && pfn_valid.

Monitoring Recommendations

Enable kdump and persistent kernel logging to capture panic traces for post-incident triage.
Alert on repeated kernel oops events from the same host, which may indicate active local exploitation attempts.
Track kernel package versions across the fleet to identify hosts still running vulnerable amdxdna builds.

How to Mitigate CVE-2026-74716

Immediate Actions Required

Apply the upstream kernel patches referenced in Kernel commit 4a19f7ab and Kernel commit 13339132, or update to a distribution kernel that incorporates the fix.
Restrict access to /dev/accel/* device nodes so only trusted users and services can open the amdxdna driver.
Unload the amdxdna module on systems that do not require AMD XDNA accelerator functionality.

Patch Information

The fix removes the offending vm_flags_mod() call and replaces the vm_insert_pages() pre-population with the handle_mm_fault() loop already used for the dma-buf import path. This aligns the VMA flags with the fault-time PFN insertion semantics and eliminates the BUG_ON trigger. See the commits at Kernel commit 4a19f7ab and Kernel commit 13339132.

Workarounds

Blacklist the amdxdna kernel module on hosts that do not use AMD NPU/AI Engine hardware.
Enforce strict permissions and ACLs on accel device nodes to limit exposure to unprivileged users.
Use container and workload isolation to prevent untrusted code from reaching the accelerator device interface.

bash

# Blacklist the amdxdna module until a patched kernel is deployed
echo "blacklist amdxdna" | sudo tee /etc/modprobe.d/blacklist-amdxdna.conf
sudo rmmod amdxdna 2>/dev/null || true
sudo update-initramfs -u