KVM nested SVM (CVE-2026-99312)
A tenant with root inside their own VM on an AMD hypervisor with nested virtualization enabled can read and write host memory, which includes the memory of every other VM on that hypervisor. A cloud provider has confirmed this on its own hosts. No supported distribution had released a fixed kernel when this page was last checked. Disable nested virtualization on AMD hypervisors now, then update the kernel when your distribution releases the fix. See Mitigation.
Summary
| Field | Detail |
|---|---|
| CVE | CVE-2026-99312 |
| Component | Linux kernel, KVM/x86, nested SVM (AMD) |
| Type | Use-after-free of a page table, reachable from inside a VM |
| Impact | Read and write access to host physical memory from inside a VM |
| Severity | High. The Linux kernel CNA scores it CVSS 3.1 8.8. See Severity ratings. |
| Privileges needed | Root inside a VM. No access to the hypervisor or the control server is needed. |
| Public exploit | None published. No exploitation in the wild has been reported. |
| Upstream fix | c7c288d178cb, in Linux 7.4-rc2 and in stable 7.3.5, 6.18.64 and 6.12.118 |
| Status checked | 19 November 2026 |
| Where it applies | AMD hypervisors with nested virtualization enabled. This does not affect Intel hypervisors or the control server. |
Am I affected?
The vulnerability is in the host kernel on your hypervisors. A hypervisor is exposed if all of the following are true:
- It has an AMD CPU. Intel hosts are not affected.
- Nested virtualization is enabled. This is the kernel's default, so a host is exposed unless it has been turned off.
- It runs an affected kernel without the fix. See Affected kernel versions.
- A VM is controlled by an untrusted party with root inside it, which is the normal situation in a hosting environment.
Check the CPU vendor. AuthenticAMD means the hypervisor has an AMD CPU:
grep -m1 vendor_id /proc/cpuinfo
Check whether nested virtualization is on. 1 or Y means enabled, 0 or N means disabled:
cat /sys/module/kvm_amd/parameters/nested
Check the running kernel:
uname -r
A hypervisor where nested virtualization was already disabled, for example as the mitigation for Januscape or Zapscape, is not exposed.
The CPU mode does not protect a hypervisor. With nested virtualization enabled, host-passthrough, host-model and the AMD CPU models all give a VM what it needs to reach the vulnerability, so changing the CPU mode of a server or hypervisor is not a mitigation. Disable nested virtualization on the hypervisor instead.
What is not affected:
- Intel hypervisors. The vulnerable path exists only in KVM's AMD code.
- The control server. It does not run VMs.
- The operating systems inside your VMs. Updating a VM's own kernel changes nothing. The fix is to the hypervisor's kernel.
What an attacker can do
The vulnerability was found by Oren Yomtov of Terra Security, who demonstrated the following:
- Reading host memory. From inside a VM at a cloud provider, about 127 GB of the host's RAM was mapped read-only and searched.
- Writing host memory. On the researcher's own AMD server, chosen bytes were written into a page of the running host kernel.
The cloud provider, DigitalOcean, has stated that it reproduced the exploit on its own hosts and treated it as critical.
Host memory holds the memory of every VM on the hypervisor, so a tenant who can read it can read other tenants' data, and one who can write it can change the host kernel. Running code on the host has not been demonstrated publicly, but write access to kernel memory is normally enough to achieve it. Treat an exposed hypervisor as one where a tenant could take full control.
No exploit code or proof of concept has been published, and there are no reports of exploitation in the wild. The published description is detailed enough that this should not be relied on for long.
No guidance for detecting an attempt has been published.
How it works
On AMD, KVM gives each VM a table that maps the VM's memory to host memory, and the processor finds that table through a pointer. While a VM runs a nested VM of its own, the processor uses a different table for the nested VM, and KVM restores the outer VM's pointer when the nested VM stops.
One path skipped that step. When a virtual CPU received an INIT (a processor reset signal) while it was running a nested VM with paging turned off, KVM returned to the outer VM without resetting its memory-management state. The outer VM was left with a pointer to a table that KVM had already freed. Once the host reuses that freed page, the VM can fill it with entries of its own choosing, and the processor follows them to any host memory the VM names.
The error was introduced in 2022 by the fix for an earlier vulnerability, CVE-2022-3344. The fix for CVE-2026-99312 resets the memory-management state on every INIT.
Affected kernel versions
| Kernel series | Status |
|---|---|
| Before 5.15.81 | Not affected |
| 5.15, from 5.15.81 | Affected. No upstream release contains the fix yet. |
| 6.1 | Affected. No upstream release contains the fix yet. |
| 6.6 | Affected. No upstream release contains the fix yet. |
| 6.12 | Affected before 6.12.118 |
| 6.18 | Affected before 6.18.64 |
| 7.3 | Affected before 7.3.5 |
| 7.4 | Fixed from 7.4-rc2 |
The 6.8 kernel of Ubuntu 24.04 and the 5.14 kernels of RHEL 9 and its rebuilds are maintained by their distributions and carry the vulnerable change. The 4.18 kernel of RHEL 8 does not.
Severity ratings
The ratings differ a great deal, because some vendors scored the issue from the kernel's description of the fix before the research was published.
| Source | Rating | Notes |
|---|---|---|
| Linux kernel CNA | High, CVSS 3.1 8.8 | Scored as a page-table use-after-free with full loss of confidentiality, integrity and availability across the VM boundary |
| SUSE | Important, CVSS 3.1 7.8 | |
| Ubuntu | Medium | |
| Red Hat | Moderate, CVSS 3.1 5.6 | Preliminary. Scored as denial of service only. |
| DigitalOcean | Treated as critical | Reproduced on its own hosts |
Do not plan around the lower ratings. The demonstrated impact is reading and writing host memory.
Is a fix available for my distribution?
Checked on 19 November 2026. The upstream fix reached the stable kernels on 17 November, and distributions had not yet released it.
| Distribution | Fix available? | Details |
|---|---|---|
| Debian 12 (bookworm) | No | linux 6.1.192-2 is vulnerable. No upstream 6.1 release contains the fix yet. |
| Debian 13 (trixie) | No | linux 6.12.117-1 is vulnerable. Upstream fixed it in 6.12.118. |
| Ubuntu 22.04 LTS | No | Ubuntu lists the kernel as "Needs evaluation". The 5.15 kernel is in the affected range. |
| Ubuntu 24.04 LTS | No | Ubuntu lists the kernel as "Needs evaluation". The 6.8 kernel is in the affected range. |
| AlmaLinux 9 and Rocky Linux 9 | No | Affected. No erratum. Red Hat lists the RHEL 9 kernel as affected with the fix deferred, so no date is known. |
| AlmaLinux 10 and Rocky Linux 10 | No | Affected. No erratum. Red Hat lists the RHEL 10 kernel as affected with the fix deferred, so no date is known. |
Until a fix is listed for your distribution, treat every AMD hypervisor with nested virtualization enabled as affected. RHEL 8 and the distributions built from it are not affected.
Sources: Debian CVE tracker, Ubuntu CVE tracker, Red Hat assessment, AlmaLinux errata, Rocky Linux errata.
Mitigation
Disabling nested virtualization on the hypervisor closes this path completely. It is the only protection available until your distribution releases a fixed kernel.
While it is off, VMs can no longer run their own VMs. Tenants who run a hypervisor, an Android emulator or Docker Desktop inside a VM will find that it stops working, so tell them before the change. Everything else in a VM is unaffected.
The setting cannot be changed while VMs are running, so it is applied by a reboot of the hypervisor. Run as root on each AMD hypervisor.
- Find any active line that already sets
nestedforkvm_amd:
grep -RHn '^[^#]*kvm_amd.*nested' /etc/modprobe.d/ /run/modprobe.d/ /usr/lib/modprobe.d/ /lib/modprobe.d/ 2>/dev/null
- If a line sets
nested=1, change it to0in the file the command printed. Replace the file name in this example with that file:
sed -i '/kvm_amd/s/nested=1/nested=0/' /etc/modprobe.d/kvm.conf
If the command printed nothing, add the setting:
echo "options kvm_amd nested=0" >> /etc/modprobe.d/kvm.conf
- Check the result. Every line printed must end in
nested=0:
modprobe -c | grep '^options kvm_amd'
- Reboot the hypervisor during a planned maintenance window:
reboot
- Confirm it is off. The value should be
0orN:
cat /sys/module/kvm_amd/parameters/nested
If it is still on after the reboot, check whether kvm_amd.nested=1 is set on the kernel command line:
grep -o 'kvm_amd.nested=[^ ]*' /proc/cmdline
If it is, remove it and reboot again.
- Debian and Ubuntu: remove
kvm_amd.nested=1fromGRUB_CMDLINE_LINUXandGRUB_CMDLINE_LINUX_DEFAULTin/etc/default/grub, then runupdate-grub. - AlmaLinux and Rocky Linux: remove it from
/etc/default/grubas well, then rungrubby --update-kernel=ALL --remove-args="kvm_amd.nested=1".
Remediation
When your distribution releases a fixed kernel, install it and reboot the hypervisor. This page will list the fixed versions as they are released.
A hypervisor is not protected by installing the kernel alone. It keeps running the old kernel until it is rebooted. After the reboot, confirm the running version with uname -r.
Keep nested virtualization disabled until the hypervisor is running a fixed kernel. If your tenants need it, remove the nested=0 setting after that and reboot once more. If they do not, leave it off.
Timeline
| Date | Event |
|---|---|
| 2022 | The fix for CVE-2022-3344 introduces the error |
| 2026-10-21 | Reported to the KVM maintainers, who write the fix |
| 2026-11-02 | Upstream fix merged |
| 2026-11-09 | Fix tagged in Linux 7.4-rc2 |
| 2026-11-17 | Stable kernels 7.3.5, 6.18.64 and 6.12.118 released with the fix |
| 2026-11-18 | CVE-2026-99312 published and the research disclosed |
| 2026-11-19 | Linux kernel CNA scores it 8.8. DigitalOcean confirms it reproduced the exploit. Ubuntu opens its record. |