Skip to content
All posts

Linux KVM CVE-2026-99312: AMD Guests Read Host Memory

November 21, 2026 · Team PlayCISO
TL;DR · the short version

CVE-2026-99312 is a flaw in KVM, the Linux kernel hypervisor, that affects only AMD hosts with nested virtualisation enabled. When KVM emulates an INIT reset while a nested guest runs, it keeps a stale page-table context, which Terra Security turned into a guest-to-host escape: on a DigitalOcean Droplet it read about 127 GB of host RAM, and on its own bare-metal AMD server it wrote chosen bytes into a live host kernel page. The Linux kernel CNA published the CVE on 18 November 2026 (CVSS 3.1 8.8); fixes are in stable 6.12.118, 6.18.64 and 7.3.5 (17 November) and 7.4-rc2. The mitigation is to set kvm_amd nested=0. DigitalOcean says it live-patched its fleet and Droplet customers need do nothing. No exploitation is reported and it is not in CISA KEV.

Put it into practice
Practise the nested-virtualisation risk decision
Work through when nested virtualisation is worth the extra attack surface on a shared host, and what to check before you allow it.
Open the lab

A virtual machine on an AMD server running Linux KVM with nested virtualisation turned on could read, and in a lab write, the host's memory. The flaw is CVE-2026-99312, published by the Linux kernel CNA on 18 November 2026 and scored CVSS 3.1 8.8. Terra Security, which reported it on 21 October, says it read about 127 GB of host RAM from a DigitalOcean Droplet and, on its own bare-metal AMD server, wrote chosen bytes into a live host kernel page. Fixed kernels shipped on 17 November. DigitalOcean says it has live-patched its fleet and Droplet customers need do nothing. Intel hosts are not affected, and no exploitation has been reported.

Who this is for: anyone who runs KVM hosts on AMD EPYC or Ryzen with untrusted or multi-tenant guests (private clouds, CI runners, hosting, VDI, labs that allow nested virtualisation), and anyone who buys VMs and wants to know what to ask their provider.

Based on the CVE.org record, the kernel.org fix, Terra Security's write-up, DigitalOcean's 19 November post, and the Red Hat, Ubuntu and Debian trackers, all read on 21 November 2026. PlayCISO is not affiliated with Terra Security or DigitalOcean. We publish no exploit details.

Confirmed vs not yet known

The kernel CNA confirms (CVE record, 18 Nov): KVM did not reset the MMU context when emulating an INIT while a nested (L2) guest ran with paging off, so the L1 guest kept a stale context. The fix commit says this "only affects nested SVM". Fixed in 6.12.118, 6.18.64, 7.3.5 and 7.4-rc2.

Terra reports: a read of about 127 GB of host RAM from a DigitalOcean Droplet, and a write into a host kernel page on its own server. DigitalOcean confirms the report came through its bug bounty programme and that it live-patched its hypervisors.

Red Hat says: exploitation requires nested virtualisation on AMD-based KVM hosts; Intel KVM and AMD hosts without nesting are not affected.

Not established: whether anyone other than the researchers used it; which other providers were exposed (Terra names none); when DigitalOcean's rollout finished; and when the 6.6, 6.1 and 5.15 kernel branches, and most distributions, will ship a fix.

What the flaw is, without the exploit

Nested virtualisation lets a guest run its own virtual machines. On AMD hardware, KVM keeps a pointer to the page tables that translate the nested guest's memory. Per the CVE record, when a nested guest receives an INIT signal (a CPU reset) while running with paging turned off, KVM skipped the step that rebuilds that context. Terra and DigitalOcean describe the result as a stale pointer to a page-table page KVM had already freed. A guest that controls what later lands in that freed page can make the CPU translate its memory accesses into host physical memory.

The CNA's own scoring notes say kvm_amd loads with nested=1 by default. Terra says the bug sat in the kernel for "just under four years" and came in with an earlier security fix; DigitalOcean says the 2022 fix for CVE-2022-3344 missed it. The CVE record traces it to a commit first released in Linux 6.1 and backported to 5.15.81 and 6.0.11.

Severity scores differ. The kernel CNA gives 8.8 (local vector, low complexity, scope changed). Red Hat's preliminary score is 7.8 (high complexity). NVD had not scored it on 21 November.

Useful next, if this was relevant to you

Showing 1 of 3: VM Escape Triage Lab

Fixed kernels and distribution status

LineStatus (21 Nov 2026)Source
MainlineFixed in 7.4-rc2 (tagged 9 Nov); commit c7c288d178cb merged 2 Novkernel.org, Terra
Stable 7.3Fixed in 7.3.5 (17 Nov)CVE record
Longterm 6.18Fixed in 6.18.64 (17 Nov)CVE record
Longterm 6.12Fixed in 6.12.118 (17 Nov)CVE record
Longterm 6.6, 6.1, 5.15Affected (5.15 from 5.15.81); no fix listed in the CVE recordCVE record
UbuntuPriority High; 22.04, 24.04, 26.04 and 26.10 kernels listed as vulnerable on 20 Nov; 20.04 not affectedUbuntu
DebianFixed only in unstable (7.3.5-1); bookworm, trixie and forky vulnerableDebian
Red HatPreliminary CVSS 7.8; mitigation published; check its CVE page for product statusRed Hat

Timeline

DateEventSource
Dec 2022Linux 6.1 ships the change that introduced the bug (also backported to 5.15.81 and 6.0.11)CVE record
21 Oct 2026Terra Security reports the flaw to the Linux KVM maintainersTerra
2 Nov 2026Fix merged to mainlineTerra
9 Nov 2026Fix first tagged in v7.4-rc2Terra, DigitalOcean
17 Nov 2026Stable 7.3.5, 6.18.64 and 6.12.118 released with the fixTerra, DigitalOcean
18 Nov 2026CVE-2026-99312 published by the Linux kernel CNA (CVSS 3.1 8.8); Terra publishes its write-upCVE.org, Terra
19 Nov 2026DigitalOcean says it live-patched its hypervisors; Ubuntu publishes its tracker entry (priority High)DigitalOcean, Ubuntu
20 Nov 2026CISA KEV catalog 2026.11.20 does not list CVE-2026-99312CISA KEV

What security leaders should do this week

  • Find your AMD KVM hosts and check nesting. On each host, cat /sys/module/kvm_amd/parameters/nested shows 1 (on) or 0 (off). Include Proxmox and OpenStack compute nodes, CI and build runners, VDI and lab hosts. Intel hosts are not affected.
  • Turn nesting off where nobody needs it. The mitigation from Terra, DigitalOcean and Red Hat is kvm_amd nested=0. Red Hat's steps: shut down running VMs, modprobe -r kvm_amd, modprobe kvm_amd nested=0, and write options kvm_amd nested=0 to a file in /etc/modprobe.d/ so it persists.
  • Patch where nesting must stay on. Move to 6.12.118, 6.18.64, 7.3.5 or later, or your distribution's fixed kernel when it ships, and reboot or live-patch. If you are on 6.6, 6.1 or 5.15, there is no upstream fix yet: keep nesting off or move branch.
  • Rank multi-tenant hosts first. The attacker needs root inside a guest. Hosts that run guests you do not control (customers, CI jobs from pull requests, student labs) come before single-owner hosts. Score it with the SSVC calculator rather than the 8.8 alone.
  • Ask your cloud providers. Patching your guest kernel does not change host exposure. Ask whether they run KVM on AMD with nesting enabled and when they patched. DigitalOcean has answered publicly; Terra says many providers it surveyed were not exposed.
  • Review your hypervisor baseline. Use the KVM Hardening Checklist to make nesting an explicit, approved exception, and log any host you cannot fix this week in your risk register with an owner and a date.

For other recent kernel flaws, see Linux Ceph CVE-2026-99304 and the Linux kernel entry in the Breach Index; for the vendor view, see AMD in the Breach Index. CVE-2026-99312 is a vulnerability in Linux KVM, not a breach of AMD or the kernel project. Practise the trade-off in the Nested Virtualization Risk Lab.

Frequently asked questions

What is CVE-2026-99312?

CVE-2026-99312 is a flaw in KVM, the hypervisor built into the Linux kernel. When KVM emulates an INIT reset while a nested guest is running on an AMD host, it can keep using a stale page-table context. Terra Security, which reported it, showed a virtual machine reading host memory and, on its own server, writing into a host kernel page. The Linux kernel CNA published the CVE on 18 November 2026 and scores it CVSS 3.1 8.8.

Does CVE-2026-99312 affect Intel hosts?

No. The fix commit says the bug only affects nested SVM, AMD's virtualisation extension, because on Intel VMX an INIT is always intercepted. Red Hat's statement says Intel-based KVM deployments are not affected.

Which kernels fix CVE-2026-99312?

Upstream stable 6.12.118, 6.18.64 and 7.3.5 (released 17 November 2026) and mainline 7.4-rc2, per the CVE record and Terra. The record lists no fix for the 6.6, 6.1 or 5.15 branches as of 21 November 2026. Distribution kernels follow their own trackers: on 20 November Ubuntu listed 22.04 to 26.10 as vulnerable, and Debian had a fix only in unstable (7.3.5-1).

How do I mitigate CVE-2026-99312 without a new kernel?

Turn off nested virtualisation on AMD KVM hosts. Terra, DigitalOcean and Red Hat all give the same mitigation: set the nested parameter of the kvm_amd module to 0. Red Hat notes running VMs must be shut down first, and the setting should be made persistent in /etc/modprobe.d so it survives a reboot.

Is CVE-2026-99312 exploited in the wild?

No exploitation has been reported. It is not in the CISA Known Exploited Vulnerabilities catalog (checked 21 November 2026). The demonstrations described by Terra and DigitalOcean were controlled tests by the researchers.

Do DigitalOcean customers need to do anything?

DigitalOcean says no. Its 19 November post says it live-patched its hypervisor kernels and tells Droplet customers "No action needed on your part". It does not say when the rollout finished or whether any customer data was read during testing.

Am I exposed if I run virtual machines in a cloud?

The flaw is in the host, not in your guest. Exposure depends on whether your provider runs KVM on AMD hosts with nested virtualisation enabled and has patched. Terra says many providers it surveyed were not vulnerable because they do not enable nested virtualisation on AMD. Ask your provider; patching your own guest kernel does not change host exposure.

Was AMD or the Linux kernel project hacked?

No. CVE-2026-99312 is a vulnerability in the Linux kernel's KVM code for AMD processors, not a breach of AMD or of the kernel project, and not a flaw in AMD hardware.

Sources

Practise the nested-virtualisation risk decision
Work through when nested virtualisation is worth the extra attack surface on a shared host, and what to check before you allow it.
Open the lab
Take this with you
What the flaw is, without the exploit · Fixed kernels and distribution status · Timeline · What security leaders should do this week · Sources

Ready to practise the decisions these articles describe?

Run your first Cyber War Room free →

Pro