NVIDIA GPU Driver EUVD-2026-90114

Requires the attacker to already have a local, authenticated low-privileged account on a system running an affected NVIDIA GPU Display Driver (GeForce, RTX/Quadro/NVS, or Tesla) on Windows or Linux - per CVSS AV:L there is no remote or network exploitation path. Impact is maximized on multi-user or multi-tenant GPU deployments (shared workstations, VDI, containerized/cloud GPU nodes) where breaking cross-process GPU-memory isolation crosses a real trust boundary. Limiting factors: local access prerequisite, PR:L (some authenticated foothold needed), and availability of the vendor driver update which closes the flaw once applied.

This is a genuine local privilege-escalation priority rather than a hollow high-CVSS number. The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H = 7.8) is internally consistent with the description: a kernel-mode flaw in the NVIDIA GPU Display Driver lets an already-authenticated local low-privileged user read memory belonging to another user's process, with full high impact across confidentiality, integrity and availability. No EPSS score, SSVC decision, KEV listing, or POC was supplied in the input, so exploitation probability and automatability cannot be quantified from the data. The local-only attack vector (AV:L) is the dominant risk-limiting factor - there is no remote or network reachability, so this is an escalation/lateral-movement primitive for an attacker who already has code running on the host, not a mass-exploitable perimeter bug. The impact is disproportionately serious on shared-GPU environments (multi-user workstations, VDI, containerized/cloud GPU nodes, ML training hosts) where cross-process/cross-tenant memory isolation is the security boundary being broken. Note CWE-862 (Missing Authorization) here describes a memory-isolation/authorization gap in the kernel driver, not a classic auth-bypass of a login mechanism, so the 'Authentication Bypass' tag overstates the class.

A low-privileged local user on a shared GPU host (e.g. a VDI seat, ML training box, or multi-tenant cloud GPU node) runs a crafted program that issues calls into the vulnerable NVIDIA kernel-mode driver to read or corrupt GPU-mapped memory owned by another user's process. By harvesting secrets from or tampering with the victim process they achieve information disclosure and, via memory corruption in kernel context, potential code execution and privilege escalation. No public exploit code is identified in the provided data, and exploitation requires prior local access to the machine.