CVE-2026-43133: Linux Kernel Privilege Escalation Bug

CVE-2026-43133 Overview

CVE-2026-43133 is a Linux kernel vulnerability in the nested SVM (nSVM) virtualization subsystem. The flaw resides in the KVM (Kernel-based Virtual Machine) module, specifically in the VMLOAD and VMSAVE instruction emulation logic for AMD SVM (Secure Virtual Machine). KVM was supposed to always reference vmcb01 for state controlled by VMSAVE/VMLOAD, following commit cc3ed80ae69f. However, the emulation path was not updated, causing KVM to operate on vmcb02 when an L2 guest issues VMSAVE or VMLOAD without L1 interception. This results in incorrect virtual machine control block (VMCB) state handling during nested virtualization.

Critical Impact
Nested guests on AMD SVM hosts can trigger incorrect VMCB state manipulation, leading to guest state corruption or unexpected hypervisor behavior in nested virtualization environments.

Affected Products

Linux kernel versions containing the original commit cc3ed80ae69f ("KVM: nSVM: always use vmcb01 to for vmsave/vmload of guest state") prior to the fix
KVM nested virtualization on AMD SVM-capable processors
Distributions shipping affected kernel versions until backported fixes are applied

Discovery Timeline

2026-05-06 - CVE-2026-43133 published to NVD
2026-05-06 - Last updated in NVD database

Technical Details for CVE-2026-43133

Vulnerability Analysis

The vulnerability exists in the KVM nested SVM implementation that handles AMD virtualization extensions. AMD SVM uses VMCB structures to store guest CPU state. In nested virtualization, KVM maintains vmcb01 for L1 guest state and vmcb02 for L2 guest state representations. The VMLOAD and VMSAVE instructions save and restore a defined subset of guest state to and from a VMCB.

A prior fix (commit cc3ed80ae69f) standardized that VMSAVE/VMLOAD-controlled fields must always be backed by vmcb01. The emulation handler responsible for emulating VMLOAD/VMSAVE when executed by L2 guests was not updated to match. When an L2 guest executes VMSAVE or VMLOAD and L1 has not configured interception of these instructions, KVM emulates the operation against the currently active VMCB, which is vmcb02, instead of vmcb01. This produces inconsistent guest state and breaks the invariant the original commit established.

Root Cause

The root cause is an incomplete refactor [CWE-1188 / kernel virtualization state desynchronization]. The emulation routine used the current VMCB pointer rather than explicitly selecting vmcb01. State written or read during emulated VMLOAD/VMSAVE diverges from the VMCB the rest of the KVM nSVM code references.

Attack Vector

Exploitation requires the ability to execute virtualization instructions inside an L2 guest on an AMD host running a vulnerable Linux KVM hypervisor with nested virtualization enabled. An L2 guest issues VMSAVE or VMLOAD, and L1 must not intercept the instruction. The emulator then operates on vmcb02 instead of vmcb01, corrupting state used by the nested virtualization logic.

No public exploit has been disclosed. The EPSS score is 0.024%, indicating a low predicted likelihood of exploitation. Code examples are not included because no verified proof-of-concept is available; refer to the upstream commits for the corrective changes.

Detection Methods for CVE-2026-43133

Indicators of Compromise

Unexpected guest state inconsistencies in nested L2 virtual machines on AMD SVM hosts
Kernel logs from KVM showing anomalous VMCB transitions or guest state divergence after VMLOAD/VMSAVE emulation
Crashes or instability in L2 guests immediately following VMSAVE or VMLOAD execution paths

Detection Strategies

Inventory Linux hosts running AMD SVM with nested virtualization enabled (kvm_amd nested=1) and confirm kernel build hashes against patched stable trees
Monitor kernel version banners and /proc/version across the fleet to identify hosts missing the upstream fix commits
Track KVM tracepoints (kvm:kvm_nested_vmrun, kvm:kvm_nested_vmexit) for unusual VMLOAD/VMSAVE emulation patterns from L2 guests

Monitoring Recommendations

Audit hypervisor hosts for kernel patch status on a recurring basis using configuration management tooling
Forward kernel logs and KVM tracepoint data to a centralized logging platform for correlation
Alert on L2 guest crashes or hypervisor warnings referencing nSVM emulation paths

How to Mitigate CVE-2026-43133

Immediate Actions Required

Apply the upstream Linux kernel patches that update the VMLOAD/VMSAVE emulation code to always use vmcb01
Track distribution security advisories and update kernels on AMD virtualization hosts as patched packages become available
Restrict the ability to launch nested guests to trusted operators until hosts are patched

Patch Information

The fix is available across multiple stable kernel branches. Refer to the upstream commits: Linux kernel commit 0004ecb798b3, commit 10063e1251c1, commit 127ccae2c185, commit 3880e331b0b3, commit c3b701500098, commit d464cf1ed900, and commit fce2fd4a2ca0. These changes correct the emulation handler to reference vmcb01 consistently.

Workarounds

Disable nested virtualization on AMD SVM hosts by loading the kvm_amd module with nested=0 until the kernel is patched
Avoid running untrusted L2 workloads on hypervisors that have not received the fix
Configure L1 hypervisors to intercept VMSAVE and VMLOAD where operationally feasible to prevent the emulation path from executing on vmcb02

bash

# Disable nested virtualization on AMD SVM hosts
sudo modprobe -r kvm_amd
echo "options kvm_amd nested=0" | sudo tee /etc/modprobe.d/kvm-amd-nested.conf
sudo modprobe kvm_amd

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