EUVD-2026-90061
GHSA-rfr4-57qm-8p8q
Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessGain code execution inside guest VM
-
DeliverySend crafted request to vGPU Manager kernel layer
-
ExploitTrigger out-of-bounds read (CWE-125)
-
ExecutionLeak host kernel memory or crash service
-
ImpactInfo disclosure, DoS, or escalation
Vulnerability Assessment
| Weakness |
CWE
ASVS 5.0
V15 - Secure Coding and Architecture Bounds checking, read limits
MITRE ATT&CK
1 technique across 1 tactic: Collection The technique list with ATT&CK links and the kill-chain view is part of the Basic plan and up. |
| Exploitation Conditions | Requires the attacker to already have code execution inside a guest VM running on a host with the NVIDIA vGPU Virtual GPU Manager (kernel-mode layer) installed - i.e., a virtualized/vGPU deployment such as VDI or cloud GPU hosting. Per CVSS PR:L, a low-privilege foothold within the guest is needed; no user interaction (UI:N) and no network access (AV:L, local to the vGPU host stack via the guest interface) are required. Limiting factors: the flaw is not reachable on bare-metal/non-vGPU installations, is not remotely exploitable over a network without a guest foothold, and is an out-of-bounds read whose escalation to code execution is described only as a possibility ('might lead to'), making information disclosure and denial of service the most reliable outcomes. |
| Impact | This is a moderate-priority local privilege/isolation flaw, not an internet-facing emergency. The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H, 7.8 HIGH) is internally consistent with the description: exploitation is LOCAL (AV:L) and requires some privilege (PR:L) - meaning an actor already executing inside a guest VM - not remote unauthenticated access. The 'RCE' and 'Buffer Overflow' tags overstate the flaw: this is an out-of-bounds READ (CWE-125), and 'RCE' is loose given AV:L (local, not network). No EPSS score, no CISA KEV listing, and no public POC are present in the provided data, so exploitation status is 'no public exploit identified at time of analysis.' The genuinely elevated concern is context, not raw score: the vGPU Virtual GPU Manager runs on the hypervisor host in multi-tenant VDI/cloud-GPU environments, and the description states a GUEST can trigger the OOB read, so the real risk is a malicious or compromised guest reading/affecting the host kernel layer that mediates other tenants. Notably the vendor set S:U (scope unchanged) despite this being guest-triggered code in a host-side manager - verify whether cross-tenant impact should push scope, but do not treat that as confirmed. Overall: real but bounded - prioritize on multi-tenant GPU hosts, deprioritize on single-tenant workstations. |
| Exploit ScenarioDerived | An attacker with the ability to run code inside a guest VM on a shared vGPU host (e.g., a low-privileged tenant in a VDI or cloud-GPU deployment) issues crafted requests to the Virtual GPU Manager kernel-mode layer, triggering an out-of-bounds read. This can leak adjacent kernel memory (information disclosure) or crash the manager (denial of service), and per NVIDIA may be chained toward code execution or privilege escalation on the host. No public proof-of-concept is available in the provided data. |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Sep 30, 2026
Out-of-bounds read (CWE-125) in the kernel-mode layer of NVIDIA's vGPU Virtual GPU Manager for Windows and Linux allows a low-privileged process inside a guest VM to force the host-side vGPU driver to read past the end of an allocated buffer (CVSS 3.1 7.8, AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H), primarily exposing host memory contents and creating crash and denial-of-service risk, with code execution and privilege escalation described by the vendor only as possible rather than demonstrated. The issue is reachable only on hosts actively running vGPU configurations such as VDI or cloud GPU hosting where an attacker already holds a foothold inside a guest; bare-metal and non-vGPU installations are not affected, and no user interaction or network reach is required. No public exploit code or confirmed active exploitation has been identified at time of analysis, and no fixed driver version appears in the available intelligence, so defenders should treat patch confirmation through NVIDIA's advisory as the priority action.
RemediationAI
No vendor-released patch version is identified in the available data at time of analysis, so the first step is to open NVIDIA's product-security bulletin 5861 (https://github.com/NVIDIA/product-security/tree/main/2026/5861, also referenced via https://nvd.nist.gov/vuln/detail/CVE-2026-47499) and apply the corrected vGPU Virtual GPU Manager host driver as soon as NVIDIA publishes it, coordinating guest and host driver compatibility since mismatched components can break vGPU passthrough. Until a fix is deployed, compensating controls should focus on removing the attacker foothold the bug depends on: do not run untrusted or multi-tenant workloads on hosts that share vGPUs, and instead pin tenants to dedicated GPUs or hardware-isolated partitions such as MIG, accepting the trade-off of lower GPU consolidation density and higher cost. Enforce least privilege inside guests so that only vetted processes can issue GPU workloads (the flaw requires only low privileges once a guest foothold exists), and monitor host and guest logs for GPU driver faults, guest crashes, or unexpected resets that could indicate attempted out-of-bounds reads; disabling vGPU entirely on a host eliminates the exposure but removes GPU virtualization capability, so it is only practical for non-vGPU-required systems.
Affected VersionsENISA
Affects your stack
Find out whether the stacks and servers you run are on an affected version.
Checking this CVE against your own stacks and servers is part of the Basic plan and up.
Technical ContextAI · show details
The flaw sits in the kernel-mode layer of the NVIDIA vGPU Virtual GPU Manager, the host-resident driver component that brokers GPU operations requested by guest VMs through the vGPU virtualization stack. The CPE data identifies the affected component as nvidia:virtual_gpu_manager with unspecified version wildcards, indicating the whole product family is flagged rather than a named release. CWE-125 (Out-of-bounds Read) means a length, index, or offset supplied or influenced by the guest is used without adequate bounds validation, causing the host kernel driver to read beyond the intended buffer. Because the read happens in privileged host kernel context on behalf of a guest, the realistic outcomes are disclosure of host-kernel or other-guest memory contents and instability that can be triggered as a denial of service; the vendor's wording that exploitation 'might lead to' code execution, privilege escalation, or data tampering indicates those stronger impacts are contingent and unproven, so analysis should weight information disclosure and DoS as the primary effects. The CVSS vector confirms local access from the guest stack (AV:L), no user interaction (UI:N), and a low-privileged requirement (PR:L), which aligns with a guest user or process that already has a code-execution foothold inside the VM rather than an unauthenticated remote attacker.
Know when Virtual Gpu Manager gets its next vulnerability
- Watch products and get an email when something new is published
- Check this CVE against your own stack and servers
- Exploit and PoC references, threat intelligence from 22 sources
No credit card · 7-day full trial
Severity by source
AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H
Guest-resident low-privilege actor (PR:L) triggers a local (AV:L) OOB read with no UI; I retain vendor S:U and high C/I/A per the described code-exec/info-disc/tamper potential.
Primary rating from Vendor (nvidia).
CVSS VectorVendor: nvidia
DescriptionCVE.org
NVIDIA vGPU Virtual GPU Manager for Windows and Linux contains a vulnerability in the kernel mode layer where a guest could cause an out-of-bounds read. A successful exploit of this vulnerability might lead to code execution, denial of service, escalation of privileges, information disclosure, and data tampering.
Articles & Coverage 3
Executive ViewAI
A high-severity memory-safety flaw in NVIDIA's vGPU Virtual GPU Manager affects only hosts actively running virtual GPU workloads, such as VDI platforms and cloud GPU hosting, where a low-privileged process already resident inside a guest VM can drive the host-side driver to read beyond an allocated buffer. The practical exposure is disclosure of host memory contents plus host crash and denial-of-service risk that would affect every tenant sharing the affected GPU host, with code execution and privilege escalation described by the vendor only as possible rather than demonstrated. Because no vendor patch is available, leadership should treat this as an unmitigated risk requiring tighter tenant-trust and configuration controls rather than a routine update cycle.
No vendor-released patch identified at time of analysis. Interim mitigations: restrict vGPU host access to trusted tenants and suspend untrusted or self-service guest provisioning, since exploitation requires an attacker to already hold a low-privileged foothold inside a guest VM; segregate or relocate sensitive workloads to bare-metal or non-vGPU hosts where feasible; apply vGPU partitioning and guest-to-host isolation hardening per NVIDIA guidance; monitor hypervisor and vGPU driver crash telemetry for indicators of out-of-bounds reads; and obtain and track NVIDIA's advisory for CVE-2026-47499 to deploy a fixed driver as soon as one is released.
Threat Intelligence Overview
NVD
CVE.org
EPSS
Vulnrichment
How this decision was reached
The engine reads each signal below and applies a fixed list of rules from top to bottom. The first rule that matches sets the recommended action – like a statement, every line is shown.
What every action level means
SSVC AssessmentENISA
Priority Score
Attack Technique
A buffer overflow occurs when a program writes more data to a memory buffer than it was allocated to hold, causing the excess data to spill into adjacent memory regions.
Information disclosure occurs when an application unintentionally exposes sensitive data that aids attackers in reconnaissance or directly compromises security.
Remote Code Execution represents the critical moment when an attacker successfully runs arbitrary code on a target system without physical access.
Denial of Service attacks render applications or systems unavailable by overwhelming resources or triggering failure conditions.
More in Virtual Gpu Manager
View allLocal privilege escalation and code execution in the NVIDIA Display Driver for Linux (GeForce, RTX/Quadro/NVS, Tesla, an
An out-of-bounds read in the kernel-mode layer of NVIDIA GPU Display Driver on Windows and Linux allows a local, authent
Local privilege escalation and potential code execution in NVIDIA GPU Display Driver for Windows and Linux arise from a
Local privilege escalation in the kernel-mode layer of the NVIDIA GPU Display Driver for Linux allows an unprivileged lo
Privilege escalation in NVIDIA's GPU Display Driver for Windows lets a local, low-privileged authenticated user bypass a
Local privilege escalation in NVIDIA Display Driver for Linux (GeForce, RTX/Quadro/NVS, Tesla, and Virtual GPU Manager b
Local privilege escalation in NVIDIA Display Driver for Windows (GeForce, RTX/Quadro/NVS, Tesla, and vGPU guest/manager
Incorrect resource transfer between spheres in the NVIDIA vGPU Virtual GPU Manager for Linux allows a locally positioned
An out-of-bounds write in the kernel-mode layer of NVIDIA vGPU Virtual GPU Manager for Linux lets a local attacker with
Local privilege escalation, code execution, denial of service, and information disclosure affect Linux hosts running NVI
Local privilege escalation in NVIDIA vGPU Virtual GPU Manager for Linux stems from an out-of-bounds read in firmware (CW
Local attackers with low privileges on a Linux host running NVIDIA vGPU Virtual GPU Manager can trigger an out-of-bounds
Exploit-DB