NVIDIA GPU driver: improper access control in the kernel mode layer allows local denial of service
Impact
An unprivileged local user can reach a kernel mode layer path that does not enforce access control properly and knock out GPU availability. Confidentiality and integrity impact are none; this is availability only. On a shared GPU node that is still real: any tenant with a GPU pod can take the device away from the other tenants on that node, and recovering a wedged GPU driver generally means a reboot and losing the running jobs. The guest driver and Virtual GPU Manager are both in the affected list.
Who can reach it
Any local low-privileged user with access to the GPU device nodes - a tenant container is enough. No privileged role needed.
What to do
Update the GPU display driver, guest driver and vGPU manager to the fixed branches in NVIDIA bulletin 2026/5861. Drain the node and reboot to reload the kernel modules.
References
Related entries
- NVIDIA GPU driver for Linux: unprivileged NULL pointer dereferences crash the driverCVE-2026-47517 · NVIDIA GPU Display Driver for Linux (NULL pointer dereference via ioctl)Medium
- NVIDIA GPU driver: divide by zero in the kernel mode layer crashes the driverCVE-2026-47534 · NVIDIA GPU Display Driver kernel mode layer (divide by zero)Medium
- NVIDIA GPU driver: uninitialized kernel memory is copied back to userspaceCVE-2026-47555 · NVIDIA GPU Display Driver kernel mode layer (uninitialized memory copied to userspace)Medium
- NVIDIA GPU driver for Linux: memory leak in error paths exhausts kernel memoryCVE-2026-47566 · NVIDIA GPU Display Driver for Linux (kernel memory leak in error paths)Medium
- NVIDIA Linux GPU driver: tenant can exhaust the DRM VMA offset address space and wedge the GPUCVE-2026-47567 · NVIDIA GPU Display Driver for Linux (kernel mode layer, DRM VMA offset allocator)Medium
- NVIDIA Linux GPU driver: unrate-limited error path lets a local user flood the kernel logCVE-2026-47568 · NVIDIA GPU Display Driver for Linux (kernel mode layer, unrate-limited error logging)Medium
This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.