Linux KVM Flaw CVE-2026-99312 Lets Guest VMs Access Host Memory on AMD Hosts
KVM is the hypervisor built into the Linux kernel. A hypervisor lets one physical computer run multiple virtual machines, or VMs, as separate guest systems. KVM provides the virtualization support that lets Linux hosts create and run those guests. It also handles boundaries between a guest and the host. In the flaw described here, KVM used nested page tables to translate a VM’s memory addresses into real memory on the host. That translation is central to keeping a guest within the memory assigned to it, rather than exposing the host’s memory or another VM’s. The bug affected a particular KVM path for nested virtualization on AMD processors, not every Linux VM host. Researchers with root access in a guest reached host memory, including about 127 GB in a DigitalOcean test. Fixed upstream kernels are available, but operators must update and reboot the host; updating only a VM’s operating system does not fix KVM.
What is KVM, and what role does it play in running virtual machines on Linux?
KVM is the hypervisor built into the Linux kernel. A hypervisor lets one physical computer run multiple virtual machines, or VMs, as separate guest systems. KVM provides the virtualization support that lets Linux hosts create and run those guests.
It also handles boundaries between a guest and the host. In the flaw described here, KVM used nested page tables to translate a VM’s memory addresses into real memory on the host. That translation is central to keeping a guest within the memory assigned to it, rather than exposing the host’s memory or another VM’s.
The bug affected a particular KVM path for nested virtualization on AMD processors, not every Linux VM host. Researchers with root access in a guest reached host memory, including about 127 GB in a DigitalOcean test. Fixed upstream kernels are available, but operators must update and reboot the host; updating only a VM’s operating system does not fix KVM.
What could an attacker inside a vulnerable virtual machine do to the host's memory?
The flaw could break the boundary between a virtual machine and its host. A guest with the required access could read host memory that it was never meant to see. On a shared server, that memory could include data belonging to other customers’ VMs.
In a DigitalOcean test, researchers running their own Droplet first saw the host’s kernel version banner, then about 127 GB of host RAM mapped read-only. In a separate test on their own bare-metal AMD server, they wrote 32 chosen bytes into a live page of the host kernel’s memory. A witness module confirmed that the bytes had changed.
The write test was not run on a shared customer server. Both demonstrations began with root access inside a VM controlled by the researchers; network access alone was not enough. The flaw affected AMD hosts with nested virtualization enabled. The provider patched its fleet, and operators of their own hosts need a fixed kernel and a reboot, or must disable nested virtualization temporarily.
How much host memory did researchers read, and what did they demonstrate in a separate write test?
The read test exposed about 127 GB of memory from the host server. The researchers ran it from a VM, or Droplet, on DigitalOcean. The host’s kernel version banner appeared first, followed by the memory, which was mapped read-only. This showed that a guest could reach far beyond the memory assigned to it.
The researchers made a different demonstration for writing. On their own bare-metal AMD server, with no other customers attached, their guest wrote 32 chosen bytes into a live page of the host kernel’s memory. A witness module loaded into the host kernel confirmed that the change had taken place. The write test was separate from the cloud read test.
Both tests started with root access inside a VM controlled by the researchers. The flaw required an AMD host with nested virtualization enabled. The provider confirmed that the exploit worked on its own hosts and patched its fleet with a live patch. Operators running their own affected hosts need to update and reboot the host kernel, or disable nested virtualization until they can do so.
Which hosts were vulnerable, and what access and settings did an attacker need?
The vulnerable systems were KVM hosts running on AMD processors with nested virtualization enabled. The vulnerable code was in KVM’s support for that AMD feature, so Intel-based KVM hosts were not affected. The researchers also found many surveyed cloud providers were not vulnerable because they had not enabled nested virtualization on AMD.
An attacker needed to control a guest VM on a vulnerable host and have root access inside that guest. The researchers began both demonstrations from that position. They did not simply attack the server over a network. On a shared host, a guest controlled by one customer could potentially expose memory that included other customers’ VMs.
The host’s settings mattered as much as the processor. Disabling nested virtualization closes this path, although it means guests can no longer run their own VMs. A fixed host kernel also closes the flaw once installed and the host is rebooted. Updating operating systems inside the guest VMs does not change the vulnerable host kernel.
What could cloud operators do to protect hosts before installing a fixed kernel?
Before a fixed host kernel is installed, the article identifies one available protection: turn nested virtualization off. This closes the vulnerable path completely. It is a host setting, not a change to make inside each virtual machine, because the flaw lives in KVM on the host.
The tradeoff is that guest VMs can no longer run VMs of their own while nested virtualization is disabled. That may affect workloads that depend on this capability. The provider took a different step for its own fleet, applying a live patch and telling Droplet customers that they did not need to act.
For operators running their own KVM hosts on AMD hardware, the lasting fix is to install a fixed host kernel and reboot the host. A guest operating system update does nothing to repair the host’s KVM code. The article notes that distribution kernels mostly did not yet have the fix at publication, so disabling nested virtualization was the available protection until a fixed kernel could be used.
What is nested virtualization, and why might someone want a virtual machine to run another virtual machine?
Nested virtualization means that a virtual machine can run another virtual machine of its own. The outer VM is running on the physical host, while the inner VM runs inside that outer guest. The article describes this capability as something some hosts allow, and identifies it as part of the vulnerable AMD KVM feature.
A common reason to use it is to test or develop virtualization software and guest environments without needing a separate physical computer. That use is general background; the article does not name particular reasons customers enable nested virtualization. Whatever the use, the host’s KVM code must manage the extra layer of virtual machines and their memory.
The flaw arose when an INIT reset stopped a nested VM but KVM did not restore the outer VM’s page-table pointer. That left the pointer referring to freed memory. Disabling nested virtualization prevents this particular path, but also prevents guests from running their own VMs. The fix resets the relevant memory-management state on every INIT.
How do nested page tables normally help keep a virtual machine's memory separate from the host's physical memory?
A VM uses its own memory addresses, while the physical server uses real memory addresses. KVM’s nested page table translates between them, mapping a guest’s memory addresses to locations in the host’s real memory. This translation helps keep the VM’s memory separate from the host and other guests’ memory.
The processor finds the table through a pointer kept for each virtual CPU. In the flaw, an INIT signal stopped a nested VM, but KVM skipped restoring the outer VM’s pointer. The pointer then referred to a page of memory that the kernel had freed. When the kernel reused that page, the guest could fill it with entries that the processor treated as the real table.
Those forged entries could name any physical page on the host, putting host memory within reach of the guest. The fix makes KVM reset the memory-management state on every INIT. Until a fixed host kernel is installed and the host rebooted, disabling nested virtualization closes this path. Guest operating-system updates do not fix the host’s page-table handling.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web