NVIDIA Display Driver CVE-2026-24191

The attacker must already have a local low-privileged interactive or code-execution foothold on a Windows system running a vulnerable NVIDIA driver (GeForce < 596.36, RTX/Quadro/NVS < 596.36, Tesla < 596.36, or a vGPU manager/guest in the affected vGPU 16.13 / 19.4 / 20.0 branches), and must be able to issue driver IOCTLs - meaning the NVIDIA display driver must be loaded and reachable from the user's session, which is the default on any workstation, server, or VDI guest using NVIDIA graphics or compute. No user interaction by another principal is required (UI:N). Exploitation is meaningfully limited by Attack Complexity High: the attacker must reliably win a TOCTOU race between the driver's validation and use of a resource, which depends on CPU scheduling, core count, and contention, and SSVC marks the issue as not automatable. Network-only attackers, sandboxed code without driver IOCTL access, and systems running only the affected vGPU guest 539.64 / 582.16 / 595.97 in a patched-host environment fall outside the practical attack surface.

CVSS 7.8 with vector AV:L/AC:H/PR:L/UI:N/S:C/C:H/I:H/A:H paints a coherent picture: local access and low privileges are required, attack complexity is High because the attacker must reliably win a race, but a successful exploit crosses a trust boundary (S:C) with full CIA impact - consistent with kernel-mode compromise. EPSS at 0.01% (1st percentile) and SSVC 'Exploitation: none, Automatable: no, Technical Impact: total' agree that no exploitation is observed and that automation is unlikely given the race condition, while still acknowledging total technical impact when it succeeds. This is not a drop-everything internet-facing emergency, but it is a meaningful local privilege escalation primitive that is highly relevant on multi-tenant systems, VDI/vGPU environments, and shared workstations where untrusted code already runs. No KEV listing and no public PoC reduce urgency, though the broad CPE footprint (consumer through datacenter) means patch coverage across an enterprise fleet is the real operational challenge.

A low-privileged local user (for example, a standard domain user on a shared workstation, a developer on a multi-tenant build server, or a guest VM user on a vGPU-backed VDI host) runs a small program that repeatedly issues a specific IOCTL or shared-memory operation to the NVIDIA driver while a second thread mutates the referenced resource between the driver's check and its use. After winning the race, the attacker gains kernel-mode code execution or privilege escalation, then leverages that to install persistence, dump credentials, or pivot. No public exploit is identified at time of analysis, and AC:H plus SSVC 'Automatable: no' indicate exploitation is non-trivial and unlikely to be turned into a click-and-own tool.