CVE-2026-63853 Overview
CVE-2026-63853 is a Linux kernel vulnerability in the AMD GPU (amdgpu) driver's Video Core Next (VCN) v4.0 encoder ring implementation. The VCN encoder and decoder rings do not support 64-bit user fence writes, yet the driver accepted command submissions containing user fences. This mismatch between hardware capability and software validation allowed local users to submit malformed command streams to the GPU. The upstream fix rejects command submissions (CS) that include user fences on affected rings by setting the no_user_fence flag for the VCN v4.0 encoder ring.
Critical Impact
Local authenticated users can trigger memory corruption or denial-of-service conditions in the kernel through crafted GPU command submissions with user fences on VCN v4.0 rings.
Affected Products
- Linux kernel builds including the amdgpu driver with VCN v4.0 support
- Systems using AMD GPUs exposing VCN v4.0 encoder and decoder rings
- Distributions shipping vulnerable kernels prior to the referenced stable patches
Discovery Timeline
- 2026-07-19 - CVE-2026-63853 published to NVD
- 2026-07-20 - Last updated in NVD database
Technical Details for CVE-2026-63853
Vulnerability Analysis
The vulnerability resides in the amdgpu kernel driver that manages AMD GPU command submission for Video Core Next hardware blocks. VCN v4.0 encoder and decoder rings lack support for 64-bit user fence writes at the hardware level. Before the fix, the driver did not mark these rings with the no_user_fence property, meaning userspace command submissions containing user fences were accepted and forwarded to hardware.
When a local user submits a crafted command stream with user fences targeting a VCN v4.0 encoder ring, the driver processes writes the hardware cannot correctly complete. This produces inconsistent kernel state, potential out-of-bounds memory writes, and reliability failures within the GPU scheduler.
Root Cause
The root cause is missing input validation in the ring initialization path. The driver failed to declare no_user_fence = true for VCN v4.0 encoder rings, so amdgpu_cs_ioctl accepted CS submissions that referenced user fence buffers on rings whose hardware cannot honor 64-bit user fence semantics. This is a driver-level configuration flaw rather than a hardware defect.
Attack Vector
Exploitation requires local access with permissions to open the DRM render node (/dev/dri/renderD*) and issue DRM_IOCTL_AMDGPU_CS calls. An attacker submits a command buffer targeting a VCN v4.0 encoder ring with a user fence descriptor. Because the attack requires low privileges but only local access, remote exploitation is not possible.
No verified public proof-of-concept code is available. See the Linux Kernel Commit Changes for the corrective patch details.
Detection Methods for CVE-2026-63853
Indicators of Compromise
- Kernel log entries referencing amdgpu GPU resets, ring hangs, or scheduler timeouts on VCN encoder or decoder rings
- Unexpected dmesg warnings involving VCN v4.0 fence completion failures
- Repeated crashes of userspace video encoding processes accessing /dev/dri/renderD*
Detection Strategies
- Audit installed kernel versions against the fixed commits 1286b6872de0, 51f694221047, and 6bdd2ed6458d from the stable tree
- Monitor process activity performing DRM_IOCTL_AMDGPU_CS submissions from unprivileged users on systems with AMD VCN v4.0 hardware
- Correlate GPU ring reset events with the submitting user identity and process lineage
Monitoring Recommendations
- Enable kernel audit rules for ioctl operations against DRM render nodes on multi-user systems
- Alert on repeated GPU driver soft-lockups or scheduler timeouts affecting VCN rings
- Track unexpected privilege boundary crossings triggered by GPU driver faults
How to Mitigate CVE-2026-63853
Immediate Actions Required
- Apply the upstream kernel patches referenced in the stable tree commits as soon as a vendor build is available
- Restrict access to DRM render nodes to trusted local users only, using group membership and udev rules
- Inventory systems using AMD hardware with VCN v4.0 blocks to prioritize patching
Patch Information
The fix sets no_user_fence for the VCN v4.0 encoder ring so amdgpu_cs_ioctl rejects submissions containing user fences. Patched commits include 1286b6872de0, 51f694221047, and 6bdd2ed6458d. The change was cherry-picked from upstream commit fd852c048b46.
Workarounds
- Limit which local accounts can access /dev/dri/renderD* devices on shared or multi-tenant hosts
- Disable VCN encoding features in userspace applications where hardware video encoding is not required
- Where feasible, blacklist the amdgpu module on systems that do not require AMD GPU acceleration
# Restrict DRM render node access to the 'video' group
sudo groupadd -f video
sudo chown root:video /dev/dri/renderD*
sudo chmod 0660 /dev/dri/renderD*
# Verify installed kernel version against fixed stable commits
uname -r
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

