News
A guest on AMD KVM can read host memory when nested virtualization is on.
CVE-2026-99312 lets a guest virtual machine on an AMD Linux KVM host with nested virtualization read host memory, which on a shared server can include other customers' virtual machines.

Impact
- What was impacted
- Linux KVM on hosts with AMD processors and nested virtualization enabled, tracked as CVE-2026-99312. From a guest they controlled, researchers read about 127 GB of a DigitalOcean server's RAM, mapped read-only, starting with the host kernel version banner. On their own bare-metal AMD server, with no other customers attached, the guest wrote 32 chosen bytes into a live page of host kernel memory. Intel KVM hosts are not affected. The vulnerable code is only in KVM's nested-virtualization support for AMD.
- Who was impacted
- Operators of their own KVM hosts on AMD hardware, especially where nested virtualization is left on. Tenants of a shared AMD KVM server where nested virtualization is enabled, because that host memory can include other customers' virtual machines. DigitalOcean told Droplet customers that no action is needed after it live-patched its fleet. Clouds the researchers surveyed that had not enabled nested virtualization on AMD were not vulnerable.
- What you need to do
- If you run the host, install a fixed host kernel and reboot. Upstream names stable kernel 7.3.5 and long-term kernels 6.18.64 and 6.12.118. Updating the operating system inside a guest does not fix this, and a new kernel does not protect the host until reboot. Until that reboot, turn nested virtualization off with the module option kvm_amd nested=0, then reboot, after warning tenants who run a virtual machine inside a virtual machine. Read the kvm_amd nested parameter. A value of 1 or Y means it is on, and the module leaves it on by default. As of this report, Debian, Ubuntu, and Red Hat Enterprise Linux 9 and 10 did not yet ship a fixed package. Red Hat Enterprise Linux 8 is listed as not affected. If you only use a DigitalOcean Droplet, the provider says no action is needed. This report has no way to detect a past attempt, and no exploit has been published.
A flaw in KVM, the hypervisor built into the Linux kernel, let researchers read about 127 GB of a cloud server's memory from a virtual machine they controlled on DigitalOcean. On a shared server, that memory can include other customers' virtual machines. The flaw is CVE-2026-99312. It affects only hosts with AMD processors and nested virtualization enabled. Intel hosts are not affected, because the vulnerable code exists only in KVM's support for nested virtualization on AMD. Both tests started with root inside a guest the researchers controlled. An attacker needs that same position: control of a guest on a vulnerable host, not only a network path to the host.
On one DigitalOcean Droplet, the read returned the host's kernel version banner and then about 127 GB of the host's RAM, mapped read-only. The write test ran on the researchers' own bare-metal AMD server, with no other customers attached. There, the guest wrote 32 chosen bytes into a live page of the host kernel's memory, and a witness module loaded into that kernel confirmed the change. Oren Yomtov of Terra Security found the flaw and reported it through the provider's bug bounty. DigitalOcean confirmed the exploit worked on its hosts and treated it as critical. The Linux KVM maintainers were told on October 21 and merged a fix on November 2. Fixed kernels then shipped upstream the day before the research went public. DigitalOcean patched its own fleet with a live patch and told Droplet customers that no action is needed on their part. VirtFusion, which is tracking the issue for operators, says there are no reports of the flaw being used outside research. No exploit code has been published, and there is no guidance for detecting a past attempt. Many cloud providers the researchers surveyed were not vulnerable because they had not enabled nested virtualization on AMD.
KVM keeps a nested page table that translates a virtual machine's addresses into the host's memory, and the processor finds that table through a pointer stored for each virtual CPU. When a host lets a virtual machine run a virtual machine of its own, an INIT signal that stops the nested virtual machine does not restore the outer virtual machine's pointer. The pointer can keep aiming at a page the kernel has already freed. After the kernel reuses that page, the guest can fill it with entries of its own, and the processor follows them as if they were the real table. Those entries can name any physical page on the host, which is how about 127 GB of RAM came into reach of a guest that was never given it. The fix makes KVM reset that memory-management state on every INIT. The researchers say the mistake came from a 2022 patch for a different nested-virtualization bug, CVE-2022-3344. That patch switched a reset path back to the outer virtual machine's control block and missed the page-table pointer. The error then sat in the kernel for just under four years. The affected code entered the 6.1 kernel in 2022 and was carried into the 5.15.81 and 6.0.11 releases, so older kernel lines and the distribution kernels built on them are in range. Scores disagree. The Linux kernel's CVE team rates the flaw 8.8, with high impact on confidentiality, integrity, and availability. Red Hat's preliminary assessment rates it Moderate at 5.6 and treats the impact as denial of service only. DigitalOcean treated it as critical for its shared cloud.
The fix lives in the host kernel. Updating the operating systems inside the virtual machines changes nothing, and a host keeps running its old kernel until it is rebooted. Upstream, the fix shipped in stable kernel 7.3.5 and in the long-term kernels 6.18.64 and 6.12.118. As of this report, distribution kernels mostly do not have the fix yet. Debian's security tracker lists the kernels in all of its releases as vulnerable, with a fix only in unstable. Ubuntu's tracker rates the flaw High and lists its kernels as vulnerable, with no fixed package released. Red Hat lists the Enterprise Linux 9 and 10 kernels as affected, with the fix deferred, and the Enterprise Linux 8 kernel as not affected. Until a fixed host kernel is installed and the host is rebooted, turn nested virtualization off. That closes this path, but virtual machines can then no longer run virtual machines of their own, so warn tenants who depend on that before the change. Nested virtualization is on by default in the kernel's kvm_amd module, according to VirtFusion's operator guide, so an unpatched host is exposed unless the setting was deliberately changed. Read the module parameter named nested. Its sysfs file is parameters/nested inside /sys/module/kvm_amd. A value of 1 or Y means it is on. If the host cannot be updated yet, set the module option kvm_amd nested=0 and reboot. The setting cannot change while virtual machines are running. After the reboot, confirm the running kernel with uname -r and read the nested setting again.
For your AI agent
Agent instructions — copy paste to Claude
You are helping me respond to CVE-2026-99312 in Linux KVM. A guest on an AMD host with nested virtualization can read host memory and, in a lab write test, change a live host-kernel page. Do not write or run an exploit or a proof of concept. Do not try to read host memory from a guest, forge a page table, or send an INIT signal to test the bug. Do not scan hosts you do not operate. Ask before writing module configuration, installing a kernel, or rebooting a host. Facts from the report: - CVE-2026-99312 affects Linux KVM only on hosts with AMD processors and nested virtualization enabled. Intel KVM hosts are not affected. The vulnerable code exists only in KVM's support for nested virtualization on AMD. - Researchers with root inside a guest they controlled read about 127 GB of a DigitalOcean server's RAM, mapped read-only. The first data back was the host kernel version banner. On a shared server, that memory can include other customers' virtual machines. Do not claim that other customers' data was shown. The report says it can be included. - A separate write test ran on the researchers' own bare-metal AMD server, with no other customers attached. The guest wrote 32 chosen bytes into a live page of host kernel memory, and a witness module in that kernel confirmed the change. - An attacker needs control of a guest on a vulnerable host, including root in that guest in the tests described. Network access to the host, with no guest, is not enough. Do not describe this as an unauthenticated remote attack. - Oren Yomtov of Terra Security reported it through the provider's bug bounty. DigitalOcean confirmed the exploit worked on its hosts, treated it as critical, live-patched its fleet, and told Droplet customers: no action needed on your part. If I only rent a Droplet and do not run the host, do not tell me to patch DigitalOcean's kernel. - Linux KVM maintainers were told on October 21 and merged a fix on November 2. Fixed kernels shipped upstream the day before the research went public: stable kernel 7.3.5 and long-term kernels 6.18.64 and 6.12.118. Do not invent other fixed version numbers. - The affected code entered the 6.1 kernel in 2022 and was carried into 5.15.81 and 6.0.11. Older lines and distribution kernels built on them are in range. - As of this report: Debian lists kernels in all releases as vulnerable, with a fix only in unstable. Ubuntu rates it High and lists its kernels as vulnerable, with no fixed package released. Red Hat lists Enterprise Linux 9 and 10 kernels as affected, with the fix deferred, and the Enterprise Linux 8 kernel as not affected. Do not invent a fixed Debian, Ubuntu, or Red Hat package version. - The Linux kernel CVE team scores it 8.8, with high impact on confidentiality, integrity, and availability. Red Hat's preliminary score is Moderate, 5.6, and treats the impact as denial of service only. Do not collapse those two scores into one. - Nested virtualization is on by default in kvm_amd, according to VirtFusion. Read the module parameter named nested. Its sysfs file is parameters/nested inside /sys/module/kvm_amd. A value of 1 or Y means it is on. Until a fixed kernel is installed and the host rebooted, set the module option kvm_amd nested=0 and reboot. The setting cannot change while virtual machines are running. Turning it off means guests can no longer run virtual machines of their own. Warn those tenants first. A newly installed kernel does nothing until reboot. Updating a guest operating system does not fix the host. - The bug is a nested page-table pointer left aimed at freed memory after an INIT stops a nested virtual machine. The fix resets that memory-management state on every INIT. The researchers trace the miss to the 2022 fix for CVE-2022-3344. Do not reconstruct the freed-page contents or the entries a guest would write. - VirtFusion says there are no reports of use outside research. No exploit code has been published. There is no guidance for detecting a past attempt. Do not invent indicators of compromise. Context: - What was impacted: Linux KVM on AMD hosts with nested virtualization, so a guest can reach host physical memory. - Who should care: people who operate those hosts, and tenants on a shared AMD KVM server where nested virtualization is enabled. DigitalOcean Droplet customers were told they do not need to act. - What to do: on hosts you operate, install a named fixed kernel and reboot, or turn nested virtualization off and reboot until that kernel is available. Do this in order. Ask before changing a host, installing a kernel, or rebooting. Prefer read-only inventory first. Do not write or run an exploit or a proof of concept. 1. Inventory - Say whether I operate the KVM host or only rent a guest. If I only use a DigitalOcean Droplet, stop after saying the provider already told those customers that no action is needed. Do not look for a kernel package inside the Droplet and call that the fix. - On each host I operate, record the CPU vendor. Flag AMD. If the CPU is Intel, say this flaw does not apply and do not change that host for this CVE. If the vendor is neither, say this report did not cover it. - On AMD hosts, if kvm_amd is loaded, read its nested parameter. The sysfs file is parameters/nested inside /sys/module/kvm_amd. Record 1, Y, 0, N, or that the file is missing. Do not load the module just to read it. - Record uname -r. Compare it only to 7.3.5, 6.18.64, and 6.12.118. If the kernel is a distribution build, say this report did not name a fixed package for Debian, Ubuntu, or Red Hat Enterprise Linux 9 and 10, and that Red Hat Enterprise Linux 8 is listed as not affected. Do not guess that a nearby version number is the fix. - Ask which tenants run a virtual machine inside a virtual machine. Turning nested virtualization off removes that. If I do not know, say so. 2. Contain - If the host is AMD, nested virtualization is on, and the running kernel is not one of the three upstream versions above, propose setting the module option kvm_amd nested=0 and rebooting. Ask before writing that option or rebooting. State that the setting cannot change while virtual machines are running, and that nested guests will lose the ability to run their own virtual machines. - Do not propose that change on a machine where I am only a guest. 3. Remediate - Plan a host kernel of 7.3.5, 6.18.64, or 6.12.118, or a distribution package only if a tracker in front of you names a fixed package. Ask before installing. Do not invent a package name. Do not propose a kernel change for Red Hat Enterprise Linux 8 for this CVE unless a tracker in front of you says that kernel is affected. - Say that the new kernel protects the host only after reboot. Ask before rebooting. Do not treat a guest operating-system update as the fix. 4. Verify - After a reboot I asked you to perform, read uname -r and the nested parameter again. - If I did not ask you to change anything, report the current values only. 5. Report back - Which machines I operate, which are AMD, whether nested virtualization is on, the running kernel, what you changed, and what is still waiting because no fixed distribution package was named. Do not invent CVEs, package versions, or signs of a past intrusion. If a check needs access I have not given you, say what you need and stop.