NVIDIA Linux GPU driver: tenant can exhaust the DRM VMA offset address space and wedge the GPU

CVSS 5.5CVE-2026-47567NVIDIA / GPU stackcurated

Impact

A local user with access to the GPU device nodes can exhaust the DRM VMA offset address space, an unbounded per-driver resource, and deny GPU service to everything else on the node. On a shared GPU node this is a cross-tenant availability problem: one pod's mapping loop takes out the other tenants' CUDA contexts, and recovery generally means resetting or rebooting the node rather than restarting one process. NVIDIA scores it 5.5 with no confidentiality or integrity impact, so the risk is uptime, not data. The record lists the Tesla datacenter branch, the vGPU guest driver and the vGPU Virtual GPU Manager as affected, so both bare-metal and virtualized GPU fleets are in scope.

Who can reach it

Local, authenticated: any tenant with a GPU pod or shell that can open the NVIDIA device nodes. No admin rights needed. Not reachable over the network.

What to do

Update the NVIDIA GPU driver to a version listed as fixed in NVIDIA security bulletin 2026/5861 (the record does not name the fixed versions, so take them from the bulletin). Replacing the driver means unloading the kernel modules, so plan a drain and reboot of each GPU node; for vGPU fleets the Virtual GPU Manager on the host and the guest drivers both need updating, which is a host maintenance window. No mitigation is documented short of restricting who can open the GPU device nodes.

References

https://nvd.nist.gov/vuln/detail/CVE-2026-47567 https://github.com/NVIDIA/product-security/tree/main/2026/5861

Related entries

All NVIDIA / GPU stack entries

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.