A use-after-free in the Linux kernel’s KVM nested virtualization code for AMD processors, tracked as CVE-2026-99312, lets code running inside a guest VM reach host physical memory. Terra Security researcher Oren Yomtov showed it against DigitalOcean: from inside a Droplet, the exploit first returned the hypervisor’s kernel version banner and then about 127 GB of host RAM. On a shared hypervisor, that memory can include other tenants’ VMs. DigitalOcean has live-patched its fleet. Anyone running their own AMD KVM hosts (Proxmox, OpenStack, oVirt, plain libvirt/QEMU, or VPS panels like VirtFusion) has to patch or disable nested virtualization.
Root cause
The bug is in how KVM emulates an INIT signal for a vCPU that is running a nested (L2) guest on AMD SVM. When INIT pulls the vCPU out of guest mode while paging is already disabled, KVM skipped resetting the vCPU’s MMU context. The MMU kept a pointer to the nested page table (NPT) root for the L2 guest, but that page had already been freed when the nested state was torn down.
Once the host allocator hands that freed page to someone else, the guest is walking a page table it does not own. If an attacker can get the page reallocated with contents they control, they can forge page table entries that map arbitrary host physical addresses into the guest. In Terra’s write-up this gave read-only access to host memory on DigitalOcean. Writes to host memory were only demonstrated on the researchers’ own bare-metal server, not on production cloud hosts.
The upstream fix, titled “KVM: x86: Unconditionally reset MMU context when emulating INIT”, does exactly what it says: KVM now forces an MMU reset on every emulated INIT, whether or not paging was already off.
Who is affected
- AMD hosts only. The bug is in the SVM nested paging path. Intel VMX hosts are not affected.
- Nested virtualization must be enabled, and it is on by default in
kvm_amd. Check withcat /sys/module/kvm_amd/parameters/nested; a1orYmeans exposed. - The attacker needs code execution inside a guest whose vCPU exposes SVM, meaning the guest can run its own nested hypervisor. That is normal for any tenant on a host that offers nested virt, or for a guest started with
-cpu hostor withsvmpassed through.
Severity is disputed. The Linux kernel CNA rates it CVSS 8.8 (high C/I/A). NVD and Red Hat’s preliminary assessments come in around 5.5โ5.6, and Red Hat’s Bugzilla entry (2575718) calls it a denial of service. A public proof of concept that reads host memory on a production cloud is a better guide to risk than either score. A PoC targeting Proxmox VE has also been posted to exploit aggregators. We have not checked how well it works. No exploitation in the wild has been reported.
Fixed versions
The fix is in these upstream stable kernels (announced on linux-cve-announce on November 18):
| Branch | Fixed in |
|---|---|
| Stable | 7.3.5 |
| LTS 6.18 | 6.18.64 |
| LTS 6.12 | 6.12.118 |
Distro status is uneven. When it was published, Debian’s tracker listed only unstable as fixed, Ubuntu showed no fixed package yet, and Red Hat had deferred fixes for at least some products. Proxmox VE ships its own kernel builds, so check its package changelog for the backport instead of going by the upstream version number.
What to do now
- Inventory AMD KVM hosts where
kvm_amdis loaded with nested enabled. That covers hypervisors, CI runners that boot VMs, and Kubernetes nodes running KubeVirt or Kata Containers on bare-metal AMD. - Patch to a kernel that carries the fix and reboot. Live-patching works too if your vendor offers a livepatch for this CVE.
- If you can’t patch, disable nested virt. Stop or migrate guests off the host, then:Any workloads that rely on nested hypervisors (Hyper-V/WSL2 inside guests, nested KVM test rigs, some sandboxing setups) will lose hardware acceleration.
1 2 3echo "options kvm_amd nested=0" > /etc/modprobe.d/kvm-amd-nested.conf modprobe -r kvm_amd && modprobe kvm_amd cat /sys/module/kvm_amd/parameters/nested # expect 0 - Stop exposing SVM to untrusted guests even after patching, unless they actually need it. Guest CPU models that hide
svmremove this whole class of nested-SVM bugs. This is the third KVM guest-to-host memory bug we have covered this autumn, after PwnEscape and the arm64 nested-virt escape. - For multi-tenant providers: treat any host that ran untrusted AMD guests with nested virt as having possibly leaked memory, including other tenants’ data and secrets held in host memory. Decide about rotating host-held credentials on that basis.
References
- DigitalOcean: Virtualization, Responsible Disclosure, and CVE-2026-99312
- Terra Security: CVE-2026-99312: KVM Guest-to-Host Escape
- Kernel CVE announcement: linux-cve-announce, 2026-11-18
- Red Hat: CVE-2026-99312 ยท Bugzilla 2575718
- VirtFusion: KVM nested SVM (CVE-2026-99312)
- OpenCVE: CVE-2026-99312