NVIDIA vGPU Manager EUVD-2026-90067

Requires an existing low-privileged local account or code-execution context on a host running the NVIDIA vGPU Virtual GPU Manager for Linux (AV:L, PR:L); the attacker must be able to reach and send crafted input to the kernel-mode layer that services vGPU requests. No user interaction is needed (UI:N) and no special non-default feature toggle is named in the description - the vulnerable code is in the standard vGPU Manager kernel path. Limiting factors: the vulnerability is NOT remotely reachable, needs prior local access, and only affects hosts that have the vGPU Virtual GPU Manager (a virtualization/VDI/cloud-GPU deployment component) installed - general GPU driver installs without vGPU are out of scope.

Moderate, locally-scoped risk. The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H, 7.8) confirms this is a LOCAL attack requiring low privileges and no user interaction - it is NOT remotely exploitable despite the 'RCE' tag in the input metadata, which is misleading given AV:L. The underlying flaw is a CWE-125 out-of-bounds READ in the vGPU Virtual GPU Manager kernel-mode layer; the vendor's claimed full C:H/I:H/A:H impact (up to code execution and data tampering) is broad for an OOB read primitive but is stated as a possible outcome, so the high impact triple is plausible under favorable exploitation. No EPSS score, no CISA KEV listing, and no public POC are present in the provided data, so there is no evidence of active or automated exploitation - this reads as a defense-in-depth patch, not an emergency. The most meaningful real-world signal is the affected component: the vGPU Virtual GPU Manager runs on the hypervisor host in multi-tenant VDI and cloud-GPU deployments, so a low-privileged guest/tenant-side foothold that reaches this code path raises a cross-tenant isolation concern that elevates it above a typical single-host local bug. Absence of EPSS/KEV/POC is not an error (this is a freshly published NVIDIA advisory), but it does mean priority should be scheduled-patch rather than drop-everything.

A low-privileged local actor on a system running the NVIDIA vGPU Virtual GPU Manager (e.g., a tenant workload or a user on the hypervisor host) issues crafted requests to the vulnerable kernel-mode layer to trigger an out-of-bounds read. The over-read can leak adjacent kernel memory (information disclosure / DoS via crash) and, per NVIDIA, may be chained toward code execution or privilege escalation. No public exploit code is identified at time of analysis, and exploitation requires an existing local foothold rather than remote access.