Critical KVM Flaw CVE-2026-99312 Allows Guest VMs to Access Host Memory on AMD Systems
A critical vulnerability in Linux's KVM hypervisor, CVE-2026-99312, allows guest virtual machines on AMD hosts with nested virtualization enabled to read host memory, potentially exposing sensitive data.

A significant security flaw has been discovered in the Linux Kernel-based Virtual Machine (KVM) hypervisor, identified as CVE-2026-99312. This vulnerability allows a guest virtual machine, particularly on AMD host processors with nested virtualization enabled, to access and read host memory. Researchers successfully demonstrated this by reading approximately 127 GB of memory from a DigitalOcean cloud server, raising concerns about the potential exposure of other tenants' data on shared infrastructure.
The vulnerability was found by Oren Yomtov, a researcher at Terra Security, who reported it through DigitalOcean's bug bounty program. The cloud provider confirmed the exploit's efficacy on its own hosts and classified it as critical. The Linux KVM maintainers were notified on October 21 and merged a fix on November 2, with patched kernels released upstream on November 19, just before the research was made public. DigitalOcean has applied a live patch to its fleet, assuring customers that no action is required on their part.
While the exploit was demonstrated on DigitalOcean's infrastructure, the core issue lies within KVM's handling of nested virtualization on AMD processors. Intel-based KVM hosts are not affected. Many cloud providers may not be vulnerable if they have not enabled nested virtualization on their AMD hardware. The exploit requires root access within the guest VM, meaning an attacker would need control of a compromised virtual machine on a vulnerable host, not just network access.
In a test environment on a bare-metal AMD server, researchers were able to write 32 specific bytes into the host kernel's memory, which was then confirmed by a witness module loaded into that kernel. This write capability, though demonstrated in a controlled setting, highlights the potential for deeper system compromise beyond just memory reading.
The flaw's origin is traced back to a security fix applied in 2022 for a different nested virtualization bug (CVE-2022-3344). This patch inadvertently altered a reset path, leaving a page-table pointer in KVM pointing to freed memory. When this memory page was later reused by the kernel, a malicious guest could manipulate it, redirecting the processor to any physical page on the host, thus enabling the large-scale memory access.
Severity ratings for CVE-2026-99312 vary. The Linux kernel's CVE team assigned it a score of 8.8 out of 10, citing high impact on confidentiality, integrity, and availability. Red Hat's preliminary assessment, however, rated it as Moderate (5.6), focusing primarily on denial-of-service potential. DigitalOcean's critical classification reflects the significant risk to its multi-tenant cloud environment.
To mitigate this vulnerability, operators of KVM hosts on AMD hardware must update their host kernel to a patched version and reboot the host. Alternatively, nested virtualization can be disabled on the host, which completely closes this attack vector but prevents VMs from running nested VMs. The fix is available in stable kernel 7.3.5 and long-term kernels 6.18.64 and 6.12.118. Distribution kernels are still catching up, with many listing their current versions as vulnerable.
There are currently no reports of this vulnerability being exploited in the wild, and no public exploit code has been released. VirtFusion, a virtualization platform vendor, is tracking the issue, but no detection guidance for past attempts is available. This incident underscores the ongoing challenges in securing complex virtualization environments, especially as new features like nested virtualization are introduced and patched.