NVIDIA vGPU Manager EUVD-2026-90062

Requires the host to be running the NVIDIA Virtual GPU Manager (vGPU plugin) for Linux and the attacker to already possess a low-privileged user account inside a guest VM served by that vGPU host. Exploitation is performed by sending a crafted RPC message containing invalid performance-state (perf-state) list size parameters to the vGPU Manager, so a vGPU/VDI/cloud-GPU multi-tenant configuration is the enabling deployment mode; single-tenant systems without vGPU virtualization are not exposed via this guest-to-host path. Limiting factors: local attack vector (no network reach), prior authentication to a guest VM (PR:L), and the presence of the vGPU Manager component - but no user interaction is needed (UI:N) and complexity is low (AC:L).

This is a genuinely notable local privilege/escalation flaw within the virtualization trust boundary, though not a broad remote threat. The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H, 7.8) reflects a low-privileged actor operating on the local system with low complexity and no user interaction, yielding full C/I/A impact. Critically, the description clarifies that the 'local' actor is a guest VM user sending a crafted RPC message to the host-side Virtual GPU Manager plugin - so the practical risk is a guest-to-host / cross-tenant compromise in vGPU/VDI/cloud-GPU deployments, which is more serious than a same-host local EoP even though S:U keeps CVSS scope unchanged. No EPSS score, KEV listing, or public POC was provided in the input; there is no SSVC decision data. NVIDIA is the reporting source and a patch is referenced (product-security bulletin 5861), so this is vendor-confirmed and fixed rather than actively-exploited. The 'RCE' tag overstates reach - the CVSS attack vector is Local (AV:L), not Network - so this should be prioritized as a virtualization-boundary EoP, not an internet-facing RCE. Overall: real priority for multi-tenant GPU hosts, low urgency for single-tenant/desktop deployments given no evidence of exploitation.

An attacker with a low-privileged foothold inside a guest VM on a host running the NVIDIA Virtual GPU Manager crafts a malicious RPC message specifying invalid performance-state list size parameters. The vGPU plugin on the host mishandles the size, causing an out-of-bounds write in host memory that the attacker leverages toward code execution or privilege escalation across the guest/host boundary. No public exploit code was identified at time of analysis, and exploitation requires an existing authenticated presence in a guest VM.