Skip to main content

A KVM Bug Lets a Nested Guest on AMD Servers Reach Host Memory, and Debian’s Stable Releases Still Lack the Fix

CVE-2026-99312 affects Linux KVM hosts with AMD processors and nested virtualisation switched on. Fixed kernels exist, distributions are behind, and one module setting closes the hole meanwhile.

KBY News Desk9 min read
Close-up of cooling fans in a server room, showcasing technology and efficiency.

A flaw in the Linux kernel’s KVM hypervisor lets a virtual machine that runs its own hypervisor reach memory that belongs to the physical host, according to the researcher who found it. The bug, CVE-2026-99312, was published on Wednesday 18 November. It only affects servers with AMD processors where nested virtualisation is switched on. Fixed kernels are out, but when we checked Debian’s tracker on Saturday 21 November, its stable releases were still marked vulnerable, so many admins will be waiting on their distribution. Until the patch arrives there is a one-line mitigation.

Most readers who run Linux hosts with KVM can stop worrying within a minute: if the host has an Intel processor, or if nested virtualisation is off, this particular bug does not apply. If it does apply, the combination of a cheap workaround and a public write-up makes it worth acting on this week.

#Who can reach it, and who cannot

The kernel’s own description of the fix says the bug “only affects nested SVM”, AMD’s virtualisation extensions, because the equivalent reset signal is always intercepted on Intel’s VMX, so the situation cannot arise there. Nested virtualisation means a virtual machine that runs a hypervisor of its own, with further virtual machines inside it. Terra Security’s researcher, Oren Yomtov, who reported the bug, says many cloud providers the company surveyed were not vulnerable because they had not enabled nested virtualisation on AMD. The Register, which reported the story on Friday 20 November, says nested virtualisation is not commonly used in the enterprise but that clouds use it for things like powering operating-system emulators.

The record published by the kernel’s CVE team goes further. Its scoring notes say the attacker needs only a non-root user able to create and run a KVM virtual machine, and that “kvm_amd nested=1 is the default”. If that is right for your kernel, an AMD host you have not touched may well have the feature on. You do not need to trust anyone’s reading of that: on the host, cat /sys/module/kvm_amd/parameters/nested prints 1 or Y when nested virtualisation is enabled and 0 or N when it is not. If the file does not exist, the kvm_amd module is not loaded, and the host is not exposed to this flaw.

#How a reset signal frees a page the CPU still trusts

Terra’s write-up explains the mechanism in plain terms, and it is worth understanding because it shows why the fix is a single line of logic. A virtual machine believes it has its own physical memory starting at address zero. The hypervisor translates those addresses into real host addresses using a table, which AMD calls the nested page table, kept in host kernel memory. The processor finds the top of that table through one pointer in a control block that KVM keeps for each virtual CPU.

When a nested virtual machine is running, the processor uses that nested machine’s pointer instead. When the nested machine stops, KVM puts the outer machine’s pointer back. Terra says one way of stopping the nested machine, a reset signal called INIT, skips that step. The outer machine then resumes with the pointer it held before, and the page it names has since been freed. The kernel’s description says the same thing in its own words: an INIT while the nested guest is running with paging disabled leaves KVM running the outer guest “with a stale MMU context”.

The freed page does not stay free. Terra says that when the virtual machine touches memory it has not used before, KVM allocates a page for a new, bottom-level table, and it can be the page the stale pointer still names. The processor then reads an ordinary page that the guest owns as a higher level of the translation table, and the guest can write whatever it likes into it. In Terra’s words, that lets the attacker forge a table that points at any host physical page. On a DigitalOcean virtual server it mapped about 127 GB of host memory read-only and found the host’s kernel version string as proof. It demonstrated the write step on its own server, where a module loaded into the host kernel confirmed that a host page had been changed. On our reading, that is the difference between a denial-of-service bug and a guest-to-host escape: reading the host’s memory shows where the real prize is, and the ability to write to it is the prize.

#Two scores, one bug

The scores do not agree, and readers should know why. The kernel’s CVE team rates the flaw 8.8 (High) on the CVSS 3.1 scale, with a low attack complexity, and describes memory disclosure, arbitrary writes to host pages and a livelock. Red Hat’s preliminary vector for the same bug is similar but rates attack complexity as high rather than low. The Register’s report gave the score as 5.5 and called it “modest”, reasoning that the bug is not a worst-case escape because it needs AMD hardware and nested virtualisation. We do not know which score the Register was quoting, and when we last looked at the National Vulnerability Database record it had not yet been assessed. Ubuntu rates it High priority.

Our view is that the lower number is a statement about how many machines are exposed, not about how bad it is on the machines that are. For a host that carries other people’s virtual machines, whether a public cloud or a shared lab, a guest that can read and write host memory is the worst thing a hypervisor bug can do. For a single-tenant host where you control every guest, the realistic risk is much lower, because an attacker would need to be inside one of your guests already.

#What is fixed, and what is not yet

The kernel record lists the fix in 6.12.118, 6.18.64 and 7.3.5 in the stable and long-term series, and in the mainline kernel from 7.4-rc2. Affected versions listed in the record include the 5.15 and 6.0 stable series and everything from 6.1 onwards until those fixed releases. According to Terra, the bug entered the kernel in 2022, when developers fixed an earlier flaw, CVE-2022-3344, and sat there for just under four years. Terra says it reported the new bug to the KVM maintainers on 21 October, that the fix reached the mainline kernel on 2 November, and that the stable releases followed on 17 November.

Distributions are behind. When we checked Debian’s security tracker on 21 November, only the unstable release (sid), on kernel 7.3.5-1, was marked fixed; bookworm on 6.1.192-2, trixie on 6.12.117-1 and forky on 7.3.4-1 were all listed as vulnerable. Ubuntu’s page listed 22.04 LTS, 24.04 LTS and 26.04 LTS as vulnerable, with 20.04 LTS not affected. Red Hat’s page for the CVE carries its usual reminder that it backports fixes rather than rebasing, so a version number alone will not tell you whether a Red Hat kernel is patched. Check your distribution’s tracker, not uname -r against the numbers above.

There is also an open question about disclosure. The Register says Terra and Yomtov published their analysis and then took it down within hours, and that it understands DigitalOcean asked for this because the company needed more time to update its Linux implementations. DigitalOcean had not replied to the Register’s requests for comment. The write-up was readable on Terra’s site when we checked on 21 November, so the details are public either way, which is a reason not to wait for a quiet maintenance window.

#What we would do

  • Find out which hosts are exposed. On every AMD KVM host, read the nested parameter as above. Intel hosts and hosts with the parameter at zero need nothing for this CVE.
  • If nested virtualisation is on and you cannot patch yet, turn it off. Terra’s advice, which the Register repeats, is to set the nested parameter of the kvm_amd module to 0, which closes this path completely. To make that persist, put options kvm_amd nested=0 in a file under /etc/modprobe.d/ and reload the module or reboot the host. The module cannot be reloaded while virtual machines are running on it, so plan it as you would any host reboot.
  • Know what you are switching off. This is our reading rather than a vendor statement: guests that rely on nested virtualisation include virtual machines that run Hyper-V or Windows virtualisation-based security, container and emulator labs, and continuous-integration runners that start their own virtual machines. Ask the owners of those workloads before you flip the switch.
  • Patch when your distribution ships the fix. Take the kernel your vendor names, then reboot the host or use the vendor’s live-patching service if it covers this CVE.
  • If you run a public or shared platform on AMD, ask your hosting or virtualisation vendor, word for word: “Do you enable nested virtualisation on AMD hosts, and have those hosts been patched for CVE-2026-99312?”

This week’s other virtualisation work is related, so it is worth planning together. Our earlier reports cover the VMware vCenter authentication bypass and the NoisyTCP flaw that can crash Linux servers remotely. A Linux host that needs one kernel update can often take them all in the same reboot.

#What we expect next

We expect distributions to ship kernels for the supported releases over the coming days, as the fixes are already in the long-term series. That is a prediction, not a promise from any vendor. What would prove us wrong is a distribution leaving the fix for a later release, which Red Hat’s page says it may do for lower-priority issues. If a cloud provider confirms that a customer could read another customer’s memory, that changes the picture from a patching task to an incident, and we will update this story.

The wider lesson is the one Terra draws from its own research: it says 3.2 per cent of kernel CVEs were introduced by the fix for another CVE, and this is one of them. That does not argue against patching. It argues for knowing which of your hosts use the unusual features, such as nested virtualisation, that turn a kernel bug from theoretical to practical, and for keeping the option to switch them off while you wait.

Evidence trail

Sources and verification

Primary documentation and external technical references used in this article.

  1. 01CVE.org: CVE-2026-99312, the Linux kernel CNA recordcve.org
  2. 02Terra Security: CVE-2026-99312, KVM Guest-to-Host Escape (Oren Yomtov, 18 November 2026)terra.security
  3. 03The Register: Flaws in VMware and KVM need attentiontheregister.com
  4. 04Debian security tracker: CVE-2026-99312security-tracker.debian.org
  5. 05Ubuntu: CVE-2026-99312ubuntu.com
  6. 06Red Hat: CVE-2026-99312access.redhat.com
  7. 07Linux kernel commit for the fix (mainline, 7.4-rc2)git.kernel.org

About this page

Credited to

Dates

Published

Corrections

No corrections are listed for this page. Some review sweeps corrected pages without naming each one; those are in the corrections log.

Reader Interaction

Comments

Add a thoughtful note on A KVM Bug Lets a Nested Guest on AMD Servers Reach Host Memory, and Debian’s Stable Releases Still Lack the Fix. Comments are checked for spam and held for moderation before appearing.

Loading comments...

Get new fixes by email.

New runbooks, configuration traps and fixes from KBY Technologies, about twice a week.