CVE-2026-99312 let researchers read about 127 GB of a cloud server’s memory from inside a virtual machine of their own on DigitalOcean.
How CVE-2026-99312 played out
Researchers at Terra Security, led by Oren Yomtov, reported a KVM (Linux kernel virtual machine) vulnerability through a provider bug bounty program that allowed a guest VM to map and read large portions of the host's RAM on AMD-based hosts with nested virtualization enabled. The provider confirmed the exploit on its own infrastructure and treated the issue as critical for its shared cloud. Fixed kernels shipped upstream on Tuesday, and the research was made public the following day.
In a test on a Droplet, the first visible data returned to the guest was the host kernel version banner, followed by roughly 127 GB of host RAM mapped read-only. A separate write test on the researchers’ own bare-metal AMD server showed the guest writing 32 chosen bytes into a live page of the host kernel's memory; a witness module in the host kernel confirmed the change. Both tests began from root access inside researcher-controlled VMs — an attacker would need equivalent control of a guest VM, not merely network access.
The technical root cause and the fix
Terra’s technical analysis describes a pointer left pointing at freed memory. KVM confines each VM with nested page tables that translate VM addresses to host physical addresses; the processor locates those tables via a pointer stored per virtual CPU. When a reset signal (INIT) stops a nested VM, KVM skipped restoring the outer VM’s page-table pointer, leaving it pointing at a page the kernel had already freed. If the kernel later reused that freed page, the guest could populate it with forged entries the processor would follow as if they were legitimate nested page-table entries, allowing the guest to name arbitrary physical pages on the host.
The upstream fix forces KVM to reset that memory-management state on every INIT. The flaw traces to a 2022 security fix (CVE-2022-3344) that switched a reset path back to the outer VM’s control block but missed the page-table pointer; the error persisted in the kernel for just under four years. The kernel team’s CVE announcement shows the fix shipped in stable kernel 7.3.5 and long-term kernels 6.18.64 and 6.12.118.

The cyber insurance questionnaire just landed. Now what?
SOC 2, HIPAA, insurance renewals - someone has to own security strategy. Nubivance provides fractional CISO leadership without the full-time salary.
Get a security leadWho is affected, and how vendors and trackers are responding
The vulnerability affects only AMD-based KVM hosts when nested virtualization is enabled; Intel-based hosts are not affected. Many cloud providers surveyed by the researchers were not vulnerable because they had not enabled nested virtualization on AMD hosts.
Distribution trackers show the fix has not broadly reached distribution kernels. Debian’s security tracker lists kernels in all its releases as vulnerable, with a fix only in unstable. Ubuntu’s CVE tracker rates the flaw 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 Enterprise Linux 8 as not affected.
Severity ratings differ: the Linux kernel’s CVE team scored the flaw 8.8/10 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. VirtFusion reports no public evidence of exploitation outside research and notes no exploit code has been published; it also tracked the issue for operators. DigitalOcean — the provider used in the published tests — patched its fleet with a live patch and told Droplet customers, "No action needed on your part."
Immediate mitigation steps for operators running AMD KVM hosts
- Update the host kernel to a fixed version (7.3.5, 6.18.64, 6.12.118 or later) and reboot the host; the host runs its old kernel until rebooted.
- If a host cannot be updated immediately, disable nested virtualization in the host kernel module: check current setting with cat /sys/module/kvm_amd/parameters/nested (1 or Y = enabled), add options kvm_amd nested=0 to the host’s module configuration, and reboot. Confirm with uname -r and by reading the nested setting again.
- Be aware the nested setting cannot change while VMs are running, and a newly installed kernel offers no protection until the host reboots.
What this means for cloud tenants, operators, and security teams
Cloud tenants who rely on shared infrastructure should note that DigitalOcean patched its fleet and advised no customer action. Tenants who require nested virtualization (running VMs inside VMs) must be warned before operators disable it as a mitigation.
Operators running their own AMD KVM hosts must prioritize kernel updates and schedule reboots or, as an interim step, disable nested virtualization — a full closure of the attack path but one that prevents guests from hosting nested VMs.
Security teams should watch distribution trackers: many distribution kernels remain vulnerable, and the only reliable protection is a patched host kernel plus reboot. Detection guidance for past exploitation does not exist, and no exploit code has been published at the time of reporting.
DigitalOcean’s live patch and Terra Security’s disclosure illustrate both the danger and the mechanics: a subtle pointer left behind by an earlier fix opened a wide window into host memory, and the remedy required a careful kernel-level reset on every nested-VM INIT. For now, the immediate questions are operational — who will reboot, who will disable nested virtualization, and whether distribution packages will catch up to the upstream fixes.




