CVE-2026-99312: AMD KVM flaw exposes host memory to guest VMs
A critical vulnerability in Linux KVM on AMD hosts allows guest VMs to read or write host memory. Patches are available for specific kernel versions.
Researchers have identified a severe security flaw in the Linux Kernel-based Virtual Machine (KVM) hypervisor that allows virtual machines to access host system memory. The vulnerability, tracked as CVE-2026-99312, specifically affects systems running on AMD processors with nested virtualization enabled. Major cloud providers like DigitalOcean have already patched their infrastructure, but self-hosted operators must take immediate action to secure their environments.
What happened
Oren Yomtov, a researcher at Terra Security, discovered the vulnerability and reported it through DigitalOcean’s bug bounty program. During testing on a DigitalOcean Droplet, the researchers were able to read approximately 127 GB of the host server’s RAM from within their own virtual machine. On shared cloud infrastructure, this memory could contain data from other customers’ virtual machines. The provider confirmed the exploit worked on its hosts and classified the issue as critical, applying live patches to its fleet without requiring customer intervention.
The Linux KVM maintainers were notified on October 21 and merged a fix into the upstream kernel on November 2. The fix was included in stable kernel 7.3.5 and long-term support kernels 6.18.64 and 6.12.118, which shipped just before the public disclosure. While DigitalOcean and other major providers mitigated the risk quickly, many distribution kernels remain vulnerable. Debian, Ubuntu, and Red Hat Enterprise Linux 9 and 10 currently list their kernels as affected, with fixes either pending or only available in unstable branches.
Severity assessments vary across the industry. The Linux kernel CVE team assigned a score of 8.8 out of 10, citing high impacts on confidentiality, integrity, and availability. In contrast, Red Hat’s preliminary assessment rated the flaw as Moderate with a score of 5.6, focusing primarily on denial-of-service risks. Despite these differences, the ability to read host memory represents a significant breach of isolation boundaries in virtualized environments.
How it works
The vulnerability stems from a logic error in how KVM handles memory management pointers during nested virtualization resets. KVM uses a nested page table to translate guest memory addresses to host physical memory. This table is located via a pointer associated with each virtual CPU. When a nested virtual machine receives an INIT reset signal, KVM is supposed to restore the outer VM’s state. However, the code fails to restore the page-table pointer, leaving it pointing to a memory page that the kernel has already freed.
If the kernel reuses that freed memory page, an attacker controlling the guest VM can fill it with forged page table entries. The processor then treats these malicious entries as valid, allowing the guest to map any physical page on the host. This mechanism enabled the researchers to read large sections of host RAM. In a separate write test on bare-metal hardware, they successfully wrote 32 chosen bytes into a live page of the host kernel’s memory, confirming the potential for integrity violations.
Interestingly, this flaw was introduced by a previous security patch. A 2022 fix for CVE-2022-3344 adjusted the reset path for nested virtualization but inadvertently missed updating the page-table pointer. This error remained in the kernel for nearly four years. Terra Security’s analysis highlights this as an example of "PatchSlop," where security fixes unintentionally introduce new vulnerabilities, noting that similar patches have historically introduced errors in 3.2% of cases.
Key details
- Affected Systems: Only Linux KVM hosts running on AMD processors with nested virtualization enabled are vulnerable. Intel-based hosts are not affected.
- Prerequisites: Exploitation requires root access inside a guest VM on the vulnerable host. Network access alone is insufficient.
- Fixed Kernels: Upstream fixes are in kernel 7.3.5, 6.18.64, and 6.12.118. Many distribution kernels, including those from Debian and Ubuntu, remain vulnerable as of publication.
- Mitigation: If patching is not immediately possible, disable nested virtualization by setting
nested=0in thekvm_amdmodule parameters and rebooting the host. - Detection: There is currently no guidance or tooling available to detect if this vulnerability has been exploited in the past.
- Exploit Status: No public exploit code has been released, and there are no reports of active exploitation in the wild outside of research contexts.
Why it matters
For software engineers and IT leads managing private clouds or bare-metal servers, this vulnerability underscores the risks of nested virtualization. While nested VMs are useful for development, testing, and certain containerized workflows, they expand the attack surface significantly. The fact that the flaw existed for four years due to a prior patch highlights the complexity of maintaining hypervisor security. Teams relying on older long-term support kernels must prioritize updates, as standard distribution packages may lag behind upstream fixes.
The discrepancy in severity ratings also reflects the varying threat models of different organizations. For a shared cloud provider, unauthorized memory access is catastrophic because it compromises tenant isolation. For a single-tenant enterprise environment, the risk might be viewed more narrowly as a potential stability or integrity issue. Regardless of the rating, the ability to read host memory breaks the fundamental trust boundary of virtualization, potentially exposing sensitive data such as encryption keys or credentials stored in RAM.
What you can do
- Check if nested virtualization is enabled on your AMD hosts by running
cat /sys/module/kvm_amd/parameters/nested. A value of 1 or Y indicates it is active. - Update your host kernel to a patched version (7.3.5, 6.18.64, or 6.12.118) if available from your distribution vendor.
- If an update is not yet available, disable nested virtualization by adding
options kvm_amd nested=0to your module configuration and rebooting the host. - Verify the mitigation by checking the kernel version with
uname -rand confirming the nested parameter is set to 0 after the reboot. - Communicate with tenants or development teams who rely on nested virtualization, as disabling it will break their ability to run VMs inside VMs.
- Monitor security trackers for your specific Linux distribution, as fixed packages for Debian, Ubuntu, and Red Hat are still in progress or limited to unstable branches.



