KVM Hypervisor Escape CVE-2026-99312: How AMD Nested Paging Flaws Leak Host Memory

A use-after-free bug in Linux KVM nested virtualization exposes up to 127 GB of host memory on AMD processors, requiring immediate host kernel updates or module reconfiguration.

Published: 2026.11.21

Visit Official Site
Host Memory Exposure
127 GB Read

Arbitrary host physical memory read tested on multi-tenant bare metal

Prerequisite Architecture
AMD + Nested VMX

Intel KVM unaffected; requires kvm_amd with nested virtualization enabled

Patch Delivery Status
Host Kernel Fix

Guest VM patching ineffective; demands host reboot or kernel livepatch

Executive Takeaway ·Hosts running AMD hardware with nested virtualization enabled face full host-memory compromise, demanding an immediate host kernel upgrade or turning off…

INIT Intercept Drops Page Table Pointers in AMD Nested Virtualization

A vulnerability in the Linux Kernel-based Virtual Machine (KVM), tracked as CVE-2026-99312, breaks guest isolation on hosts powered by AMD processors. The flaw allows an attacker with root privileges inside a guest virtual machine to bypass memory boundaries and access host physical memory. During validation on a multi-tenant cloud environment hosted on DigitalOcean, security researchers at Terra Security mapped approximately 127 GB of host RAM into an unprivileged guest. The read sweep exposed kernel version banners and adjacent tenant memory spaces, followed by verified 32-byte arbitrary writes to host memory pages on dedicated test hardware.

The failure originates inside KVM’s implementation of AMD Nested Paging (NPT) within the kvm_amd kernel module. KVM uses nested page tables to translate guest virtual addresses into host physical addresses. When a host enables nested virtualization, a L1 guest can spawn its own L2 nested instances, requiring the hypervisor to swap page-table pointers during state transitions.

When an INIT reset signal terminates an L2 nested instance, KVM fails to restore the outer L1 guest’s page-table pointer correctly. This oversight leaves the CPU referencing a page-table address that the host kernel has already deallocated. Once the host kernel reallocates that memory page for routine operations, the L1 guest crafts arbitrary translation entries inside the reused structure. The AMD hardware virtualization unit treats these forged entries as legitimate translations, mapping host physical frames directly into the guest address space.

CVE-2026-99312 Hypervisor Boundary Escape Sequence

Memory state corruption during AMD nested guest INIT reset

01 · Step 01
1. L2 VM Reset Trigger

L1 guest issues INIT signal to reset inner nested L2 virtual machine.

02 · Step 02
2. Pointer Restoration Omission

KVM clears L2 state but fails to rebind L1 nested page-table pointer.

03 · Step 03
3. Host Memory Deallocation

Host kernel frees original page-table frame and reallocates memory.

04 · Step 04
4. Forged Address Mapping

Guest writes arbitrary PTE entries to map host physical RAM directly.

This vulnerability was introduced by a 2022 security commit designed to patch CVE-2022-3344. That earlier fix restored virtual machine control blocks during reset sequences but omitted the corresponding page-table pointer refresh. The defective code persisted across upstream kernel releases for roughly four years, entering Linux 6.1 and backporting to 5.15.81 and 6.0.11.

Intel implementations of KVM remain unaffected. The vulnerable execution path resides exclusively inside AMD-specific nested virtualization logic. Furthermore, guest-level operating system patches provide zero protection. Because the hypervisor manages address virtualization, remediation depends entirely on host-level kernel updates or disabling the nested virtualization module parameter.

Specification Matrix and Remediation Telemetry

Mitigating CVE-2026-99312 requires evaluating deployment hardware, hypervisor configurations, and patch distribution channels. Infrastructure teams operating private clouds or running bare-metal instances must distinguish between hypervisor patching requirements and tenant-level configurations.

Operational VectorVulnerable ProfilePatched / Mitigated Profile
CPU ArchitectureAMD EPYC / Ryzen (SVM Enabled)Intel Xeon / Core (VMX) or Unaffected AMD
Module Parameterkvm_amd nested=1 (Kernel Default)kvm_amd nested=0
Upstream KernelLinux 6.1 through 6.12.117 / 6.18.63Linux 6.12.118, 6.18.64, 7.3.5+
Enterprise Linux (RHEL)RHEL 9, RHEL 10 (Kernel Update Pending)RHEL 8 (Not Affected)
Guest VM IsolationStandard KVM Isolation (Compromised)Host-Enforced Hard Memory Boundaries
Remediation TargetHost Node KernelGuest OS (No Action Effective)
Service InterruptionHost Reboot (or Ksplice / Livepatch)Zero Downtime if Module Unloaded at Boot

Enterprise distributions reflect varying vulnerability assessments. The Linux kernel CVE assignment team assigned the flaw a CVSS score of 8.8, classifying it as High severity with full confidentiality, integrity, and availability impacts. Red Hat assigned a Moderate severity rating of 5.6, categorizing early risks as denial-of-service.

Immediate Module Disable vs. Kernel Upgrade

Evaluating downtime, guest features, and patching velocity

Kernel Livepatch or Reboot
  • + Preserves nested virtualization for CI/CD and sandbox environments
  • + Maintains complete multi-tenant memory integrity across host fleet
  • + Permanently closes use-after-free pointer condition upstream
Emergency Module Deactivation (nested=0)
  • – Breaks tenant workloads reliant on nested hypervisors (Docker in VM, WSL2)
  • – Requires coordination across bare-metal compute instances
  • – Leaves fleet vulnerable if configuration slips occur during reboots

Because downstream enterprise distribution pipelines exhibit release lags, Debian, Ubuntu, and Red Hat distributions have listed repository updates as pending or unstable. Public cloud providers operating shared AMD infrastructure with nested virtualization active have largely applied custom hypervisor livepatches to avert customer-facing hypervisor reboots.

Operational Remediation and Hypervisor Hardening Protocols

Remediating CVE-2026-99312 requires infrastructure operators to execute a structured validation pipeline across bare-metal hypervisor pools. Because hypervisor memory leaks leave no post-incident audit trail in guest logs, organizations must assume any exposed host running untrusted multi-tenant workloads requires immediate isolation.

Fleet-Wide Hypervisor Neutralization Workflow

Verification and remediation sequence for AMD KVM nodes

01
1. Audit Configuration Step 01

Verify CPU architecture and check /sys/module/kvm_amd/parameters/nested.

02
2. Disable Nested Virtualization Step 02

Write 0 to module parameter or deploy modprobe.d drop-in override.

03
3. Upstream Kernel Deployment Step 03

Deploy kernel 6.12.118+, 6.18.64+, or vendor-specific livepatch packages.

04
4. Host Node Reboot Step 04

Reboot host physical nodes to enforce new hypervisor binary in memory.

Protocol 1: Live Fleet Inspection and Temporary Module Neutralization

Before scheduling node maintenance windows, security teams must audit compute clusters to identify exposed hosts. Running the check reveals whether nested virtualization is currently active:

# Verify AMD virtualization module status
cat /sys/module/kvm_amd/parameters/nested

A return value of 1 or Y indicates the hypervisor actively supports nested instances. If upstream patches cannot be deployed immediately, operators should disable nested virtualization across running nodes where supported, and enforce persistent overrides:

# Write persistent override to modprobe configurations
echo "options kvm_amd nested=0" > /etc/modprobe.d/kvm-amd.conf

# Reload module configuration during next maintenance cycle
modprobe -r kvm_amd && modprobe kvm_amd

Disabling this parameter stops an L1 VM from initializing an L2 hypervisor, neutralizing the initialization routine that triggers the unmapped pointer reuse.

Protocol 2: Kernel Upgrades and Long-Term Multi-Tenant Isolation

For organizations providing container sandboxes, automated continuous integration runners, or nested staging environments, turning off nested virtualization breaks core platform functionality. These environments require updating the host operating system to Linux 6.12.118, 6.18.64, 7.3.5, or distribution-equivalent backports.

  • Host Fleet Verification: Ensure cloud orchestrators do not place untrusted workloads on unpatched AMD compute nodes. If running heterogeneous clusters comprising Intel and AMD nodes, route workloads requiring nested virtualization to Intel pools until AMD host kernels are fully patched.
  • Tenant Communication: If nested virtualization must be deactivated dynamically, inform users running emulation, Android development tools, or nested Docker instances that hardware acceleration will remain disabled until the underlying hypervisor reboot completes.
  • Supply Chain and Kernel Regression Auditing: CVE-2026-99312 marks another occurrence where maintenance patches introduced critical use-after-free conditions. Infrastructure teams maintaining custom hypervisor forks should incorporate explicit nested page-table initialization tests into their continuous delivery pipelines, preventing state-management regressions during upstream kernel merges.

Related Benchmark Reports

View All in Category →
Weekly Briefing

Weekly Tech & Business Data Briefing

Verified software analysis, practical gotchas, and essential supply-chain updates delivered weekly.

Unsubscribe with 1 click anytime. Zero spam.

* We may earn an affiliate commission from links in this report, at no extra cost to you and with zero impact on our benchmark data.