NVIDIA GPU Driver CVE-2026-47595

Requires the NVIDIA GPU Display Driver for Linux to be installed and the affected kernel-mode layer loaded, plus an already-authenticated local unprivileged user account on the host (PR:L, AV:L) - there is no remote or network-facing attack surface and no user interaction is needed. The core precondition is the driver's failure to preserve read-only permissions on a memory region, allowing an unprivileged write. Limiting factors: attacker must already have local code-execution as some user, exploitation is Linux-only (the CPE set and Linux tag confirm no Windows exposure for this CVE), and only systems actually running the vulnerable driver branch (GeForce / RTX-Quadro-NVS / Tesla / Guest Driver) are affected. Real-world impact concentrates on shared/multi-tenant GPU environments (cloud GPU, VDI, ML/render clusters) where crossing the unprivileged-to-kernel boundary is meaningful.

This is a genuine but not emergency-grade local privilege-escalation issue. The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H, 7.8 HIGH) is internally consistent with the description: exploitation requires local access and an existing unprivileged account (PR:L), no user interaction, and yields full C/I/A impact on the host. No EPSS score, CISA KEV listing, or public POC was provided, so exploitation status is 'no public exploit identified at time of analysis' - this is a vendor-self-reported (NVIDIA) flaw with a patch available via product-security bulletin 5861, not an actively exploited one. The single notable signal conflict is the 'RCE' tag: the CVSS attack vector is AV:L (local), so the 'code execution' outcome is local kernel-mode code execution / privilege escalation, NOT remote code execution - the RCE tag overstates reach. Treat this as a solid patch-in-normal-cycle item, prioritized upward on multi-tenant GPU hosts (VDI, cloud GPU, shared ML/render farms) where an unprivileged local user boundary actually matters, and lower on single-user workstations where the attacker already owning a local account greatly reduces incremental risk.

A local unprivileged user on a Linux host with the NVIDIA GPU driver installed issues crafted requests to the kernel-mode driver that cause it to write to a memory region whose read-only permissions were not preserved, corrupting kernel-controlled data. By steering this write the attacker escalates to kernel/root privileges or destabilizes the system. No public exploit code was identified in the provided data, and exploitation needs local access plus an existing account (PR:L), so this is a post-foothold privilege-escalation primitive rather than an initial-access vector.