ServicesBlogPricingContactContact Now
Cloud3 min read

Linux KVM CVE-2026-99312: Guest VMs Access Host Memory on AMD

Linux KVM CVE-2026-99312: Guest VMs Access Host Memory on AMD

Introduction

A newly disclosed vulnerability in the Linux Kernel-based Virtual Machine (KVM) subsystem, tracked as CVE-2026-99312, permits a malicious or compromised guest virtual machine to read large portions of host physical memory. Researchers demonstrated the flaw by extracting approximately 127 GB of memory from a cloud server hosted on DigitalOcean, highlighting the risk of cross-tenant data exposure in multi-tenant environments. The vulnerability is limited to hosts running AMD processors with nested virtualization enabled, and patched kernel versions have already been shipped.

Technical Details of the Vulnerability

KVM is the hypervisor framework built directly into the Linux kernel, widely used by cloud providers to isolate customer workloads. The flaw resides in the handling of nested virtualization — a feature that allows a guest VM to run its own hypervisor and child VMs. On affected AMD hardware, the KVM module fails to properly enforce memory isolation boundaries when nested virtualization is active, enabling a guest to bypass the hypervisor's memory management controls and access host physical pages that may contain data from other VMs or the host kernel itself. The exact mechanism has not been publicly detailed beyond the source summary, but the impact is clear: a guest can read memory it should never see.

Impact and Scope

The vulnerability affects only systems meeting two conditions: the host CPU must be an AMD processor, and nested virtualization must be enabled in the KVM configuration. Intel-based hosts are not impacted. Cloud providers that offer nested virtualization as a feature — often used for CI/CD pipelines, security research, or running hypervisors inside VMs — are the primary targets. In a shared infrastructure, the 127 GB memory read demonstrated on DigitalOcean could encompass encryption keys, credentials, application data, or live memory of co-located customer VMs. While the source does not specify whether write access or code execution is possible, read access alone constitutes a severe breach of confidentiality and violates the fundamental isolation guarantee of virtualization.

Mitigation and Remediation

Fixed kernel versions have been released by upstream Linux kernel maintainers and downstream distributors. Administrators should take the following steps immediately:

  • Apply the latest kernel security updates from your distribution (e.g., Ubuntu, Debian, RHEL, Fedora, SUSE, or cloud-provider kernels).
  • If patching cannot be applied immediately, disable nested virtualization on AMD hosts by setting the kernel parameter kvm_amd.nested=0 or via the module parameter nested=0 for the kvm_amd module, then reboot.
  • Verify that nested virtualization is disabled by checking cat /sys/module/kvm_amd/parameters/nested — it should return N or 0.
  • Monitor vendor advisories for any additional microcode or firmware updates that may complement the kernel fix.
  • Audit cloud provider status pages for confirmation that underlying host infrastructure has been patched.

Organizations running on-premises KVM clusters on AMD hardware should prioritize patching hypervisor nodes, especially those offering nested virtualization to internal tenants.

Grivyonx Expert Analysis

This vulnerability underscores a recurring theme in cloud security: hardware-assisted virtualization extensions are complex, and the interplay between nested paging, IOMMU, and vendor-specific implementations creates a broad attack surface. AMD's nested virtualization implementation (SVM) has historically received less scrutiny than Intel's VT-x, making it a logical target for researchers. The fact that a single guest could read 127 GB suggests the flaw may involve a missing TLB flush, incorrect ASID handling, or a race in the nested page table walk — all areas where subtle logic errors can collapse isolation. Defenders should treat nested virtualization as a high-risk feature and disable it unless strictly required. Grivyonx's AI-driven DevSecOps platform helps teams continuously validate hypervisor configurations and detect drift from hardened baselines across hybrid cloud estates.

Conclusion

CVE-2026-99312 is a potent reminder that the trust boundary between guest and host remains a critical battleground. While the vulnerability's scope is narrow — AMD hosts with nested virtualization — its impact on multi-tenant clouds is severe. Patching is straightforward and available; the greater challenge is ensuring rapid deployment across sprawling infrastructure. Organizations should also reevaluate whether nested virtualization is necessary for their workloads and disable it by default. Grivyonx's cloud security automation services can streamline the detection and remediation of such configuration risks at scale.

  • #KVM
  • #CVE-2026-99312
  • #virtualization
  • #AMD
  • #cloud security
  • #hypervisor
Gourav Rajput

Gourav Rajput

Founder of Grivyonx Technologies