NVIDIA vGPU Manager EUVD-2026-90071

Requires the host to be running the NVIDIA Virtual GPU Manager (vGPU plugin) with at least one guest VM assigned, and the attacker must already hold an unprivileged local user account inside such a guest VM to issue the crafted RPC to the host. This is inherently limited to vGPU/VDI/cloud-GPU deployments (not standard bare-metal GPU driver installs), requires no user interaction on the host, and is not remotely reachable over the network - the 'local' vector is exercised from within a guest across the virtualization RPC channel to the host.

This is a genuine but moderate-priority issue whose real-world risk is defined by its attack surface rather than by mass exploitability. The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U) confirms local access and low privileges are required - there is no network-facing exposure - but the critical nuance in the description is that the 'local' attacker is a GUEST VM USER sending a crafted RPC to the vGPU Manager on the HOST. In a multi-tenant vGPU/VDI/cloud-GPU deployment this is effectively a guest-to-host boundary crossing, which raises the practical severity above what a bare 'local privilege escalation' label suggests. The vendor scored S:U (unchanged scope) and C:L/I:H/A:H, yielding 7.3; a case could be made for scope change given the guest→host crossing, but the vendor's I:H/A:H with only C:L (an OOB write that corrupts/crashes but leaks little) is internally consistent. No EPSS score, KEV listing, or public PoC is present in the supplied data, so there is no evidence of active or widespread exploitation - this is a vendor-discovered (NVIDIA-reported) hardening fix, not a signal of in-the-wild activity. Bottom line: not a drop-everything emergency, but it should be patched on any host that runs untrusted or multi-tenant guest VMs, where the guest→host escape potential makes it materially more important than a single-tenant driver bug.

An attacker who already controls an unprivileged user account inside a guest VM on a vGPU-enabled host sends a specially crafted RPC call to the Virtual GPU Manager plugin running on the hypervisor host. The malformed RPC triggers an out-of-bounds write in the host-side vGPU plugin, corrupting host memory to escalate privileges, tamper with data, or crash the service (DoS). No public exploit code is identified in the provided data, and the local attack vector plus required guest foothold limit this to attackers who already have tenancy on the affected host.