NVIDIA Display Driver CVE-2026-24196

Attacker must have local interactive or programmatic access to the host running a vulnerable NVIDIA Linux driver and must hold low-level privileges sufficient to open NVIDIA device nodes (typically /dev/nvidia0, /dev/nvidiactl, /dev/nvidia-uvm) and issue ioctls - this is normally granted to any user in the 'video' group or any process able to use CUDA/OpenGL/Vulkan. No user interaction is required (UI:N). Exploitation is meaningfully limited by the local-only attack vector (AV:L) - there is no remote network path - and by the requirement that the host run an affected driver branch (R535 prior to 535.309.01, R580 prior to 580.159.03, R595 prior to 595.71.05, or specific vGPU Guest builds 535.288.01/580.126.09/595.58.03). The condition is therefore concrete: any Linux system with a vulnerable NVIDIA driver where untrusted users or workloads can execute code, with vGPU guests, GPU cloud instances, and shared HPC/VDI hosts being the realistic exposure surface.

This is a moderate-priority issue rather than an urgent emergency. The CVSS vector (AV:L/AC:L/PR:L/UI:N) requires local access with low privileges, which constrains the attack surface to users or processes already on the host. EPSS is 0.01% (1st percentile), CISA SSVC marks Exploitation as 'none' and Automatable as 'no' with only partial Technical Impact, and no public exploit identified at time of analysis - all signals converge that real-world exploitation is unlikely in the short term. The headline 7.1 score is driven by C:H/A:H, but the absence of integrity impact (I:N) and the local-only vector mean this is not in the same tier as remote pre-auth RCE. The most material risk is in multi-tenant environments - shared HPC clusters, GPU cloud nodes, VDI hosts, and vGPU guests - where a low-privileged tenant could read memory belonging to other tenants or crash the shared GPU stack. On single-user workstations, the practical risk is lower.

On a shared GPU host - for example a multi-tenant HPC node, a VDI server, or a vGPU guest VM - a low-privileged local user invokes a crafted ioctl or GPU API call against the NVIDIA kernel driver, causing it to read past the bounds of a kernel buffer. The attacker harvests leaked memory contents to obtain sensitive data from other tenants or kernel structures, or repeatedly triggers the read at an unmapped address to crash the driver and deny GPU service to all users on the host. No public exploit identified at time of analysis, and SSVC marks the issue as not automatable.