Skip to content

Vulnerabilities & Patches

Researchers demonstrate Linux KVM flaw reading 127 GB of host memory

Shared cloud customers on AMD hosts with nested virtualization enabled are exposed until patched and rebooted, The Hacker News reported.

https://breach-signal.com/vulnerabilities/linux-kvm-flaw-reads-127-gb-host-memory/

Share

Key facts

  • CVE-2026-99312 lets a guest VM read about 127 GB of host memory on an AMD host with nested virtualization on.
  • Researchers demonstrated the flaw from inside their own virtual machine on DigitalOcean, a shared server whose memory can include other customers' virtual machines.
  • Customers on shared cloud infrastructure with AMD hosts and nested virtualization enabled are exposed until the host kernel is updated and the host rebooted.
  • DigitalOcean verified that the attack succeeded against its own infrastructure and rated the vulnerability as critical.

The Hacker News reported on November 20, 2026, that researchers at Terra Security demonstrated CVE-2026-99312 from inside a DigitalOcean virtual machine running on a shared server, reading about 127 GB of host memory on an AMD host with nested virtualization enabled (a guest running its own virtual machine). DigitalOcean operates shared cloud hosts whose memory can include other customers’ virtual machines. Oren Yomtov of Terra Security discovered the vulnerability and reported it through DigitalOcean’s bug bounty program. DigitalOcean verified that the attack succeeded against its own infrastructure and rated the vulnerability as critical.

The flaw exists only in the AMD nested-virtualization support for kernel-based virtual machine (KVM) within the Linux kernel, so Intel-based KVM hosts are not affected. Exploiting the issue requires the attacker to control a guest instance on the vulnerable host; mere network access by itself is insufficient to exploit the flaw. Both tests described in the research began with root access inside a virtual machine controlled by the researchers, who operated the guest used for the demonstration.

A number of cloud providers examined by the researchers were not exposed because they had not activated nested virtualization on AMD processors. Customers on shared cloud infrastructure with AMD hosts and nested virtualization enabled are exposed until the host kernel is updated and the host rebooted, after which the patched kernel takes effect.

The read test on DigitalOcean first returned the host’s kernel version banner, then about 127 GB of the host’s RAM, mapped read-only. On the researchers’ own physical AMD server, with no additional customers present, the guest wrote 32 bytes of its choosing into an active page of host kernel memory. A witness module inserted into the kernel confirmed the modification, keeping the controlled write test separate from the read test on DigitalOcean.

VirtFusion, a virtualization platform vendor tracking the issue, reports no use of the flaw outside research. No exploit code has been made public, and there is no guidance for identifying a prior intrusion.

The Linux kernel CVE team gives the vulnerability an 8.8 out of 10 rating, marking confidentiality, integrity, and availability as highly impacted. Red Hat’s preliminary security assessment labels the flaw Moderate at 5.6 and describes the effect solely as denial of service.

DigitalOcean patched its own fleet with a live patch rather than asking customers to act. The provider told Droplet customers, “No action needed on your part.”

Linux KVM maintainers were told about the flaw on October 21 and merged a fix on November 2. Upstream shipped fixed kernels on Tuesday, including stable kernel 7.3.5 and long-term kernels 6.18.64 and 6.12.118. The fixed kernels arrived on Tuesday, one day before the research became public on November 20.

Debian flags the kernels across every release as vulnerable, with a patch available only in unstable. Ubuntu classifies the flaw as High and marks its kernels as vulnerable, though no fixed package has been released yet. Red Hat lists Enterprise Linux 9 and 10 kernels as affected with the fix deferred, and Enterprise Linux 8 as not affected.

Older kernel lines and distribution kernels built on those releases remain exposed. Administrators operating their own KVM hosts on AMD hardware should either update the host kernel or disable nested virtualization.

Updating the operating systems inside virtual machines changes nothing because the vulnerable code exists only in KVM’s AMD nested-virtualization support on the host. A host continues to run its previous kernel until it restarts. Until the host reboots into a fixed kernel, the only protection is turning nested virtualization off, which closes the path completely.

Turning nested virtualization off prevents virtual machines from running virtual machines of their own. The kvm_amd module enables nested virtualization by default, meaning an unpatched host remains at risk unless someone deliberately switched the setting off. VirtFusion’s operator guide confirms the module ships with nested virtualization enabled by default.

Operators can check the setting by reading the nested parameter for the kvm_amd module, where a value of 1 or Y means nested virtualization is on. If patching must wait, they can add options kvm_amd nested=0 to the module configuration file on the host and then reboot.

The setting cannot change while virtual machines are running, so any change requires a reboot before the new setting takes effect. After the system reboots, administrators should verify the patch using uname -r and by checking the nested parameter again.

Before turning nested virtualization off, operators should warn tenants who depend on it because the change stops guest VMs from running their own VMs. The vulnerable code was introduced in kernel version 6.1 during 2022 and later included in releases 5.15.81 and 6.0.11.

The researchers say a 2022 patch for CVE-2022-3344 created the flaw by missing the page-table pointer when switching a reset path to the outer VM’s control block. The bug remained in the kernel for slightly less than four years before the fix was merged.

KVM keeps a page-table pointer for each virtual CPU to translate guest addresses into host memory as part of its memory-management machinery. When an INIT reset halts a nested VM, KVM fails to restore the outer VM’s pointer, leaving it directed at memory that has been freed and available for reuse by the kernel. After the kernel recycled the freed page, the guest populated it with its own entries, and the processor followed the forged entries as the real table, letting them name any physical page on the host. The patch forces KVM to clear that memory-management state whenever an INIT occurs.

This vulnerability is the most recent in a string of KVM isolation problems disclosed during the current year. Earlier failures include a 16-year-old flaw in KVM’s shadow memory management affecting Intel and AMD systems. The PwnEscape vulnerability in Linux networking allowed a guest to read host memory inside Vercel’s live sandbox environments in October.

Terra Security’s earlier PatchSlop work found that fixes introduced 3.2 percent of kernel vulnerabilities. Terra Security published a technical analysis on Wednesday about the KVM freed-memory pointer bug that explains the underlying flaw.

Was this helpful?

Join the discussion

Please read our Comment Policy before commenting. First-time comments are held for moderation. By posting a comment you agree to our Terms of Use and Privacy Policy, which explain how your name, email address and comment are used.

Your email address will not be published. Required fields are marked *