CVE-2026-63851

What it actually does

The flaw allows command submission (CS) with user fences on VCN encoder/decoder rings that lack support for 64-bit user fence writes, bypassing intended input validation and enabling invalid memory operations from userspace.

Why it matters to the business

Local users can trigger kernel-level memory corruption or crashes on systems with affected AMD GPUs, leading to denial of service or potential privilege escalation on workstations and servers.

How an attack unfolds

An attacker with local access crafts a command submission buffer containing a user fence for a VCN 4.0.5 encode ring, submits it through the amdgpu driver interface, and causes the kernel to perform an unsupported 64-bit fence write that corrupts memory or crashes the system.

How to fix it

Update the Linux kernel to a version containing the commit that sets no_user_fence on VCN v4.0.5 encode rings and rejects unsupported user-fence submissions; apply the change via distribution security updates.

How to confirm the fix worked

Confirm the patch is present by checking for the drm/amdgpu/vcn no_user_fence logic in the amdgpu driver source or by verifying that CS submissions with user fences on VCN 4.0.5 rings are now rejected with an error.

Typical attack path

1. Obtain local shell on a Linux system with vulnerable amdgpu driver
2. Map GPU rings and identify VCN encoder/decoder rings
3. Construct a CS submission packet that includes a 64-bit user fence
4. Submit the packet via the DRM ioctl interface
5. Trigger kernel path that performs the unsupported fence write

How to detect it

Review kernel logs for amdgpu fence-related warnings or oops on VCN rings; examine audit logs for unusual DRM ioctl submissions from non-graphics processes.

What your security team should do

• Prioritize kernel updates on all systems with AMD VCN 4.0.5 GPUs
• Restrict unprivileged local access where feasible until patches are applied
• Monitor for amdgpu-related kernel oops or crashes as a temporary indicator