Threat intel5 min read
Linux KVM flaw lets a guest VM reach host memory on AMD servers
CVE-2026-99312 lets a virtual machine read and write the host's memory on AMD-based KVM hosts with nested virtualization on. Fixed kernels are out, but most distributions have not shipped them yet.
4Tify Team

A bug in KVM, the hypervisor built into the Linux kernel, lets code running inside a virtual machine read and write the physical memory of the host underneath it. It is tracked as CVE-2026-99312 and was published on 18 November 2026. Only AMD hosts with nested virtualization enabled are affected, and on those hosts nested virtualization is switched on by default.
What happened
In a demonstration on an ordinary DigitalOcean droplet, a guest VM mapped about 127 GB of the host's RAM read-only. On a separate AMD server under the researchers' own control, the same technique was used to write chosen bytes into a live host kernel page. On a shared cloud host, that memory can hold other customers' virtual machines.
The issue was reported to the KVM maintainers on 21 October. The fix was merged into the mainline kernel on 2 November and shipped in stable kernels 7.3.5, 6.18.64 and 6.12.118 on 17 November, one day before the CVE went public.
The Linux kernel CNA rates it 8.8 (High). Red Hat lists it as Important with a draft score of 7.8, and Ubuntu gives it high priority. No public exploit code had been released by 19 November, there are no reports of use in attacks, and the flaw is not in CISA's Known Exploited Vulnerabilities catalog. Full details are in the NVD entry.
Who is affected
- Hardware: AMD hosts running KVM (the
kvm_amdmodule). Intel hosts are not affected, because Intel's virtualization always intercepts the INIT signal involved. - Setting: nested virtualization on, which is the
kvm_amddefault. - Kernels: the bug arrived with a 2022 patch for CVE-2022-3344 and sat in the kernel for just under four years. It is present from 6.1 onward and in the 5.15 and 6.0 lines from 5.15.81 and 6.0.11. The CVE record lists fixes for 7.3, 6.18 and 6.12 only, so the 6.1 and 6.6 long-term lines have no fix yet.
- Distributions (as of 21 November): Debian 12 and 13 are listed as vulnerable, with a fix only in unstable. Ubuntu 22.04 and 24.04 show a fix still needed. Red Hat lists RHEL 9 and 10 as affected and RHEL 8 as not affected.
The fix lives in the host kernel. Updating the operating systems inside your VMs does not help.
How it works
KVM keeps each VM inside its own view of memory using a nested page table, which the processor finds through a pointer held for every virtual CPU. When a guest runs a VM of its own and an INIT reset hits it while paging is off, KVM skipped resetting that pointer. It kept pointing at a page the host kernel had already freed.
Once the kernel reused that page, the guest could fill it with page table entries of its choosing. The processor then followed those entries to any physical page on the host. The fix forces KVM to reset its memory management state on every INIT.
What to do now
- Check exposure. On each AMD hypervisor, run
cat /sys/module/kvm_amd/parameters/nested. A value of1orYmeans nested virtualization is on. - Patch where you can. Install a kernel that includes the fix (7.3.5, 6.18.64, 6.12.118 or a distribution build that names CVE-2026-99312), reboot, and confirm with
uname -r. A host keeps running the old kernel until it restarts. - Otherwise, turn nested virtualization off. Add
options kvm_amd nested=0to a file in/etc/modprobe.d/and reboot, or reload the module once all VMs are stopped. Check thenestedvalue again afterwards. Tenants who run their own hypervisors or other nested VMs will lose that ability, so tell them first. - Watch your distribution's tracker for the backported kernel and plan a maintenance window for the reboot.
Kernel flaws like this one are easy to miss on hypervisors that rarely get rebooted. Our vulnerability scanning service checks host kernel versions and module settings against new CVEs such as this one, so an exposed AMD host shows up before a tenant finds it.