On November 18, 2026, the Linux kernel published CVE-2026-99312, a flaw in KVM, the hypervisor built into Linux, that breaks the wall between a virtual machine and its host. On AMD hosts with nested virtualization enabled, a user with root inside a guest can make KVM use a page table that has already been freed, and through it read and write the host’s physical memory. Host memory holds the kernel and every other virtual machine on the server, so on shared infrastructure this is a cross-tenant flaw. Oren Yomtov of Terra Security showed it on a DigitalOcean Droplet, where his exploit mapped about 127 GB of host memory read-only from inside the guest. On his own AMD server he also wrote to a live host kernel page. Google’s Stefan Teodorescu and Chris Mason reported the same bug independently. DigitalOcean reproduced it on its own hosts, treated it as critical and live-patched its fleet. Its Droplet customers do not need to do anything. The fix landed in stable kernels 6.12.118, 6.18.64 and 7.3.5 on November 17. The 5.15, 6.1 and 6.6 kernel lines had no fix on November 21, and most distributions had not shipped one. Intel hosts are not affected. No attacks have been reported, the flaw is not in CISA’s Known Exploited Vulnerabilities catalog, and no exploit has been published. CVE CVSS EPSS CISA KEV Fixed in CVE-2026-99312 8.8 (CVSS 3.1), Linux kernel CNA 0.2% No Linux 6.12.118, 6.18.64 and 7.3.5, mainline 7.4-rc2. No upstream fix for 5.15, 6.1 or 6.6 on November 21 CVSS and EPSS from NVD and FIRST, exploitation status from the CISA Known Exploited Vulnerabilities catalog, checked November 21, 2026. What is confirmed Point Status Guest root can read and write host memory and hang the host Linux kernel CNA, Red Hat Only AMD hosts with nested virtualization on (kvm_amd nested=1, the default) Kernel fix commit, Red Hat Intel hosts not affected, because VMX always intercepts INIT Kernel fix commit Introduced in kernel 6.1 in 2022, and backported to 5.15.81 and 6.0.11 Linux kernel CVE announcement Fixed in 6.12.118, 6.18.64 and 7.3.5 (November 17) and mainline 7.4-rc2 Linux kernel CVE announcement No upstream fix for 5.15, 6.1 or 6.6 on November 21 Linux kernel CVE announcement Demonstrated: about 127 GB of host memory read on a DigitalOcean Droplet, 32 bytes written into a host kernel page on a lab server Terra Security, DigitalOcean DigitalOcean fleet live-patched, no customer action DigitalOcean Public exploit None CISA KEV Not listed on November 21 How the flaw works Nested virtualization lets a virtual machine run its own hypervisor and guests. KVM tracks separate page tables for this: the outer guest’s view of memory, and the inner guest’s. AMD’s nested paging tables decide which physical memory each guest can reach. When KVM emulates an INIT signal, the processor reset that starts a CPU, while the inner guest is running with paging turned off, it resets the CPU state but keeps the old memory management context. The outer guest then carries on with a nested paging root that points at a page KVM has already freed. If the guest then sets up its own nested guest, turns off paging and the INIT intercept, and sends itself an INIT, it can steer KVM onto that freed page and fill it with page tables of its choosing. Those tables can map any physical memory on the host into the guest. The bug came from a 2022 fix for an earlier flaw, CVE-2022-3344, which is why it reaches back to kernel 6.1. The new fix makes KVM always reset the memory context when it emulates INIT. Intel hosts are safe because Intel’s virtualization always intercepts INIT, so the path never runs. Practical exploit effort The kernel scores this 8.8 and Red Hat and SUSE 7.8, because it needs local access with privileges. In practice the bar depends on who controls the guests. The attacker needs root in a virtual machine on an AMD host with nested virtualization on. On a public cloud or VPS provider, anyone can rent such a machine, so the precondition costs a credit card. In a private environment, it means an attacker has already compromised a guest and become root. The technique is demonstrated, the code is not public. Terra Security mapped about 127 GB of host memory on a Droplet, read-only, and checked it by finding the host kernel’s version banner. On its own server it wrote 32 bytes into a live host kernel page, confirmed by a test module. Neither Terra nor DigitalOcean has said what else was visible in memory or how long the exploit took. No code has been published, and no one has shown running code on the host, although write access to host memory is usually enough to get there. Rebuilding it is skilled kernel work. The trigger needs a guest that runs its own nested guest with specific settings, then sends an INIT at the right moment, then builds page tables in freed memory. The fix commit and Terra’s write-up explain the mechanism, so a capable kernel exploit developer could reproduce the read in days. Turning host memory writes into reliable code execution takes longer. Nested virtualization being on is the key condition. It is the default in the kvm_amd module, but Terra says many cloud providers it checked had turned it off on AMD. A guest can tell whether its CPU offers AMD’s virtualization (svm in /proc/cpuinfo), which is how an attacker would pick targets. Failure is noisy on the host. The kernel CNA describes a livelock in nested page fault handling. A failed attempt is more likely to hang a CPU or crash the host than to stay silent. Our assessment of the effort, as of November 21: Public cloud and VPS hosts on AMD Private virtualization (Proxmox, OpenStack, oVirt) on AMD Intel hosts or AMD with nested off Public exploit No No Not affected Attacker needs A rented VM with root, which is trivial An already compromised guest with root Not reachable What it gives Read and write of host memory, including other customers’ VMs Read and write of host memory, so every VM on the host Nothing Effort Skilled kernel work for read access, more for code execution The same, after a first compromise None Loudest signal Host hangs or crashes, nested page fault storms Host hangs or crashes None For hosting providers on AMD hardware with nested virtualization on, this is a critical flaw: any customer can try it. For organizations that run their own KVM on AMD, it is a serious second-stage flaw that turns one compromised VM into the whole host. For most attackers today it needs an exploit that has not been published. Who is affected, in priority order Hosting, VPS and cloud providers running KVM on AMD hosts with nested virtualization exposed to customers. Private clouds and virtualization platforms on AMD: Proxmox VE, OpenStack, oVirt and OpenShift Virtualization, especially where untrusted users or teams get root in VMs. Development, CI and lab hosts on AMD that run nested hypervisors, Android emulators or Docker Desktop inside VMs. Customers of providers that run KVM on AMD, who should check their provider’s statement. DigitalOcean has patched. On AWS, nested virtualization is offered only on Intel instance types. On Google Cloud, nested virtualization on AMD is limited to the N4D series, and Google had not published a bulletin by November 21. Azure runs on Hyper-V, not KVM. Updating a guest’s kernel does nothing. The fix is in the host kernel. Affected and fixed versions Kernel line Affected Fixed 7.3 Below 7.3.5 7.3.5 6.18 Below 6.18.64 6.18.64 6.12 Below 6.12.118 6.12.118 6.6 Yes No upstream fix on November 21 6.1 Yes No upstream fix on November 21 5.15 From 5.15.81 No upstream fix on November 21 Mainline 7.4-rc2 Distribution status on November 21, 2026: Distribution Status Debian Fixed only in unstable (7.3.5-1). Bookworm, trixie and forky affected, no DSA yet Ubuntu Priority High. linux needs a fix on 22.04, 24.04 and later releases, including the cloud kernels. 20.04 and older are not affected. No USN yet Red Hat Important, CVSS 7.8. RHEL 9 and 10 and OpenShift CoreOS affected, RHEL 7 and 8 not affected. No RHSA yet. Red Hat’s mitigation is kvm_amd nested=0 SUSE Important. Fixed in SLES 15 SP7, SLES 16, SL Micro 6 and Leap 16. SLES 15 SP4 to SP6 and SLE Micro 5.3 to 5.5 affected AlmaLinux and Rocky Linux No erratum yet Proxmox VE No advisory. Proxmox kernel 7.3 builds from 7.3.5 carry the fix DigitalOcean Fleet live-patched, no customer action Distributions backport fixes into their own kernel versions, so check your vendor’s advisory rather than comparing version numbers with the upstream table. How to check a host These commands only read state. Run them on each KVM host: grep -m1 vendor_id /proc/cpuinfo lsmod | grep -w kvm_amd cat /sys/module/kvm_amd/parameters/nested cat /proc/cmdline | tr ' ' '\n' | grep -i nested uname -r A host is exposed if the CPU is AuthenticAMD, kvm_amd is loaded, nested shows 1 or Y, and the kernel is not a fixed build from your distribution. To see which guests have AMD virtualization exposed to them, check the CPU model each VM uses. With libvirt: virsh list --name | while read vm; do [ -n "$vm" ] && echo "$vm: $(virsh dumpxml "$vm" | grep -cE "feature policy='require' name='svm'|mode='host-passthrough'")"; done A count above zero means the guest can see svm and run nested guests. On Proxmox, VMs with CPU type host pass it through. Mitigation and remediation Install your distribution’s fixed kernel and reboot, or apply a live patch if your vendor offers one. Upstream, that means 6.12.118, 6.18.64, 7.3.5 or later. Where no fixed kernel exists yet, turn off nested virtualization on AMD hosts. Stop or migrate the VMs on the host, then: modprobe -r kvm_amd && modprobe kvm_amd nested=0 echo "options kvm_amd nested=0" > /etc/modprobe.d/kvm-amd-nested.conf Also remove any kvm_amd.nested=1 from the kernel command line. This breaks nested hypervisors, Android emulators and Docker Desktop inside VMs, so warn the teams that use them. Red Hat gives this as its mitigation. Changing the guest CPU model alone is not enough. 3. Hide AMD virtualization from guests that do not need it, by not passing the svm CPU feature through. This limits exposure where nested must stay on for some VMs. 4. Keep untrusted workloads off shared AMD hosts with nested on until they are patched, and separate tenants or teams by host where you can. 5. Intel hosts need no action for this flaw. Free SOC trialWant this watched across your environment without building the hunts yourself? Our 24x7 SOC monitors your endpoints, servers and cloud during a free trial.Start a free trialThreat hunting No attacks have been reported and there are no indicators of compromise. The exploit runs inside a guest and acts on host memory, so it leaves few traces in normal logs. Hunting focuses on finding exposed hosts, on host instability, and on root activity inside guests that have no reason to run nested virtualization. Example KQL for Microsoft Defender and Microsoft Sentinel The KQL below is example code for Microsoft Defender for Endpoint advanced hunting and Microsoft Sentinel. Adapt table and field names to your environment and test each query before relying on it. Hunt 1: KVM hosts and their kernels // Example: Defender for Endpoint, Linux devices running QEMU/KVM, with their OS version DeviceProcessEvents | where Timestamp > ago(7d) | where FileName in~ ("qemu-system-x86_64", "qemu-kvm", "kvm") | summarize VMProcesses=dcount(ProcessId), LastSeen=max(Timestamp) by DeviceId, DeviceName | join kind=leftouter (DeviceInfo | summarize arg_max(Timestamp, OSPlatform, OSVersion, OSDistribution) by DeviceId) on DeviceId | project DeviceName, OSDistribution, OSVersion, VMProcesses, LastSeen | order by DeviceName asc Check each host’s CPU vendor and nested setting with the commands above. Hunt 2: host hangs and nested paging errors // Example: Sentinel Syslog from KVM hosts, soft lockups, RCU stalls and KVM errors Syslog | where TimeGenerated > ago(14d) | where SyslogMessage has_any ("soft lockup", "rcu_sched self-detected stall", "rcu: INFO", "hung_task") or (SyslogMessage has "kvm" and SyslogMessage has_any ("BUG", "WARNING", "general protection", "npf")) | summarize Events=count(), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated), Sample=any(SyslogMessage) by Computer | order by Events desc The kernel describes a livelock in nested page fault handling as one effect of the bug. A KVM host that starts locking up after a tenant’s VM begins running nested guests is worth investigating. Hunt 3: nested virtualization started inside Linux guests // Example: Defender for Endpoint on Linux guests, processes that load KVM or start a hypervisor DeviceProcessEvents | where Timestamp > ago(14d) | where (FileName in~ ("modprobe", "insmod") and ProcessCommandLine has_any ("kvm_amd", "kvm")) or FileName in~ ("qemu-system-x86_64", "qemu-kvm", "cloud-hypervisor", "firecracker") | project Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine, InitiatingProcessFileName | order by Timestamp desc Run this against your virtual machines, not the hosts. A server VM that suddenly loads KVM or starts its own hypervisor, with no build, test or emulator role, is unusual. A custom kernel module loaded in a guest is a stronger signal, since a working exploit is likely to need one. Hunt 4: kernel modules loaded inside guests // Example: Defender for Endpoint on Linux, kernel modules loaded from unusual paths DeviceProcessEvents | where Timestamp > ago(14d) | where FileName in~ ("insmod", "modprobe") | where ProcessCommandLine has ".ko" and not(ProcessCommandLine has "/lib/modules/") | project Timestamp, DeviceName, AccountName, ProcessCommandLine, InitiatingProcessCommandLine | order by Timestamp desc Responding to a suspected compromise Keep the evidence: host kernel logs, the VM’s disk and, if possible, a memory image of the suspect guest before shutting it down. Isolate the suspect VM and move other VMs off the host. Patch or set nested=0 on the host and reboot it. Treat the host and every VM on it as exposed. Memory read access means secrets in other VMs’ memory, such as keys, tokens and passwords, may have been read. Rotate them, starting with the most sensitive workloads. Rebuild the host if host memory writes or code execution cannot be ruled out. Timeline 2022: A fix for CVE-2022-3344 lands in kernel 6.1 and is backported to 5.15.81 and 6.0.11, introducing the bug. September 25, 2026: CVE-2026-99312 is reserved. October 21, 2026: Terra Security reports the flaw to the KVM maintainers, and Google’s Sean Christopherson posts the fix. DigitalOcean is also told in October. November 2, 2026: The fix is merged into mainline, released in 7.4-rc2 on November 9. November 17, 2026: Stable kernels 6.12.118, 6.18.64 and 7.3.5 ship the fix. November 18, 2026: The Linux kernel publishes CVE-2026-99312. Terra Security publishes its write-up, takes it down within hours, and later restores it. November 19, 2026: DigitalOcean says it reproduced the flaw and live-patched its fleet. Ubuntu rates it High. The kernel CNA scores it 8.8. November 21, 2026: No fix for the 5.15, 6.1 and 6.6 lines, most distributions have no fixed package, and the CVE is not in CISA KEV. Frequently Asked Questions Is CVE-2026-99312 being exploited? No attacks have been reported, and it is not in the CISA KEV catalog. Researchers demonstrated it on a DigitalOcean Droplet, and DigitalOcean has patched its fleet. Which hosts are affected? Linux KVM hosts on AMD processors with nested virtualization enabled, which is the default, running kernel 6.1 or later, or the 5.15.81 and 6.0.11 backports, without the fix. Intel hosts are not affected. Which kernel fixes CVE-2026-99312? Upstream, 6.12.118, 6.18.64, 7.3.5 and 7.4-rc2. The 5.15, 6.1 and 6.6 lines had no fix on November 21. Use your distribution’s fixed kernel when it ships, and reboot. How do we work around it? Set kvm_amd nested=0, reload the module or reboot, and remove kvm_amd.nested=1 from the kernel command line. This stops nested hypervisors and emulators inside VMs from working. Do we need to patch our virtual machines? No. The fix is in the host kernel. Updating a guest’s kernel does not help. Are we affected on AWS, Azure or Google Cloud? AWS offers nested virtualization only on Intel instance types, Azure uses Hyper-V rather than KVM, and Google Cloud allows nested virtualization on AMD only on the N4D series. None had published a statement on this CVE by November 21. DigitalOcean has patched. Is Proxmox VE affected? Proxmox runs KVM, so AMD hosts with nested virtualization on are affected. Proxmox had no advisory on November 21. Its kernel 7.3 builds from 7.3.5 carry the fix, and setting nested=0 is the workaround for older kernels. What could an attacker read? Any host memory, including the host kernel and other VMs on the same server, so keys, passwords and data in those VMs’ memory. How ThreatDefence is responding ThreatDefence has added CVE-2026-99312 to our vulnerability intelligence. Across all monitored environments, our SOC is: identifying Linux KVM hosts on AMD processors and checking their kernel and nested virtualization settings watching for host lockups and kernel errors on KVM hosts, and for guests that load KVM or kernel modules without a reason to helping customers apply fixed kernels or the nested=0 workaround and plan the reboots If you want help finding exposed virtualization hosts or investigating unusual guest activity, contact us. Read more about our managed detection and response.