NVIDIA vGPU EUVD-2026-90066

Exploitation requires local access to a host running the NVIDIA vGPU Virtual GPU Manager for Linux with existing low-level privileges (PR:L) - for example a low-privileged local account or a guest/tenant able to reach the vGPU Manager firmware interface. The vulnerability resides in the firmware and is triggered by causing an out-of-bounds read; no user interaction is needed (UI:N) and attack complexity is low (AC:L). Limiting factors: no network vector (AV:L rules out remote unauthenticated exploitation despite the 'RCE' tag), the target must actually be deployed as a vGPU/VDI/cloud-GPU host running the affected Virtual GPU Manager, and the attacker must already possess a local foothold. Only Linux deployments of the Virtual GPU Manager are affected.

This is a genuine but not top-tier priority: a real local privilege-escalation class flaw whose CVSS of 7.8 is honest for what it is, but which is not remotely exploitable. The CVSS vector CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H confirms LOCAL attack vector requiring EXISTING low privileges (PR:L) - this is not a network-facing RCE despite the 'RCE' tag. No EPSS score, no CISA KEV listing, and no public POC were provided, so exploitation status is 'no public exploit identified at time of analysis'; the vulnerability was reported by NVIDIA itself (coordinated disclosure) rather than discovered in active attacks. No SSVC decision data was supplied. The key tension is between the alarming tag set ('RCE', 'Buffer Overflow') and the actual vector: the flaw is a firmware out-of-bounds READ (CWE-125), requires local authenticated access to the vGPU Manager host, and any code execution would be a local privilege escalation, not remote takeover. The elevated real-world concern is contextual: the Virtual GPU Manager sits on the hypervisor/host in multi-tenant vGPU, VDI, and cloud-GPU deployments, so a low-privileged guest or local user reading out of bounds in host firmware could threaten tenant isolation - but this requires an already-established foothold. Net: a valid patch-soon issue for vGPU operators, not an emergency internet-facing exposure.

A low-privileged local user or a tenant with guest-level access on a system running the NVIDIA vGPU Virtual GPU Manager for Linux issues crafted requests to the vulnerable firmware interface, triggering an out-of-bounds read. The leaked adjacent memory contents are used to defeat mitigations and, in the most severe case, escalate to code execution in the host/kernel context, potentially crossing the tenant isolation boundary in a VDI or cloud-GPU environment. No public exploit code was identified at the time of analysis, and exploitation requires pre-existing local access (AV:L/PR:L).