NVIDIA GPU Driver CVE-2026-47598

Requires: (1) the affected NVIDIA open-source kernel module variant installed and loaded on a Linux host - the flaw is specific to the open-source module's event-delivery path and does not affect Windows; (2) an authenticated local user with access to the NVIDIA GPU device nodes (PR:L); and (3) winning a race between asynchronous event delivery and file close (AC:H), which is non-deterministic and typically demands repeated attempts. Limiting factors: no remote vector, no exploitation without local access, and the race condition substantially reduces reliability. Risk is elevated specifically in multi-tenant GPU deployments (vGPU Manager, Guest Driver, VDI, cloud GPU) where low-privileged tenants share the host kernel.

CVSS 3.1 base score is 7.0 (HIGH) with vector AV:L/AC:H/PR:L/UI:N/S:U/C:H/I:H/A:H, and I concur with this scoring: the CWE-416 use-after-free is triggered by a race between asynchronous event delivery and file close, so AC:H correctly reflects the non-deterministic timing window an attacker must win, and AV:L/PR:L correctly capture that exploitation requires an already-authenticated local user with access to the NVIDIA device nodes. No EPSS score, no CISA KEV listing, and no public POC are present in the supplied intelligence, so all exploitation signals point to no observed or probable exploitation at time of analysis. This is a genuine local privilege-escalation primitive rather than a remotely wormable flaw - the 'RCE' tag overstates reach because the vector is strictly local. Real priority is MODERATE: the AC:H race condition and PR:L requirement meaningfully raise the exploitation bar for a single-user workstation, but the impact profile (full C/I/A on the kernel) makes it materially more attractive in multi-tenant contexts (vGPU, VDI, cloud-GPU) where a low-privileged tenant escaping to kernel code execution breaks isolation. No SSVC decision data was provided; treat automatable=no as the likely posture given the race dependency.

An authenticated low-privileged local user on a Linux host running the affected NVIDIA open-source kernel module opens the GPU device, registers for asynchronous events, then repeatedly races event delivery against closing the file descriptor to trigger a use-after-free on the freed event structure. By grooming the kernel heap to reclaim the freed object with attacker-controlled data, they escalate to kernel-level code execution. No public exploit code is identified at time of analysis, and the AC:H race window means reliable exploitation typically requires many attempts.