NVIDIA vGPU EUVD-2026-90063

Requires authenticated local access (PR:L) to a Linux host running the NVIDIA vGPU Virtual GPU Manager - i.e., a low-privileged local user or a guest context on the hypervisor - able to reach the kernel-mode vGPU layer to submit the malformed request that triggers the out-of-bounds read. No user interaction is needed (UI:N). Exploitation is LIMITED to systems actually deploying NVIDIA vGPU on Linux (VDI/cloud-GPU virtualization hosts); it is not remotely reachable (AV:L) and cannot be triggered by an unauthenticated network attacker. The most impactful scenario is a multi-tenant vGPU deployment where the OOB read crosses tenant isolation boundaries.

This is a genuine but not emergency-grade 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) makes clear the attack is LOCAL and requires existing low privileges - not remote. The 'RCE' tag is therefore misleading: code execution here is local privilege-driven, not network-exploitable. No EPSS score, CISA KEV entry, or public POC is present in the input, so there is no public exploit identified at time of analysis and no evidence of active exploitation. The realistic threat is a low-privileged tenant/guest on a vGPU host (VDI or cloud-GPU multi-tenant environment) leveraging a kernel-mode out-of-bounds read to leak memory, crash the host, or escalate - a meaningful cross-tenant isolation concern in shared-GPU deployments, but requiring a foothold first. Priority is MEDIUM: high CVSS reflects the C:H/I:H/A:H kernel impact, but local-only access and absence of exploitation signals keep real-world urgency moderate. Note the description lists 'code execution ... escalation of privileges, information disclosure, and data tampering' as possible outcomes of what is fundamentally an OOB READ (CWE-125) - write/tamper/exec outcomes from a pure read primitive are plausible but uncertain; verify exploitability scope with NVIDIA bulletin 5861.

A low-privileged user or guest VM on a Linux hypervisor running NVIDIA vGPU Manager issues crafted requests to the kernel-mode vGPU layer, triggering an out-of-bounds read. This leaks adjacent kernel memory (potentially another tenant's data) or crashes the host GPU stack (DoS), and under favorable conditions is chained toward privilege escalation. No public exploit code is identified; exploitation requires an existing local foothold on the vGPU host.