What happened
Oren Yomtov, a researcher at the security firm Terra Security, found a flaw in KVM, the hypervisor built directly into the Linux kernel. KVM powers a large share of cloud and VPS servers, as well as many in-house virtualization servers.
He reported it through DigitalOcean’s bug bounty program. From inside their own virtual machine (a Droplet), the researchers were able to read about 127 GB of the host server’s memory, the provider said. The Linux KVM maintainers were told on October 21, 2026 and merged a fix on November 2, 2026. The flaw was added to the NVD database on November 18, 2026 as CVE-2026-99312.
No one has reported exploitation outside research so far. VirtFusion, a virtualization platform vendor tracking the issue, knows of no attacks, and according to The Hacker News no exploit code has been published. The flaw is not in the CISA KEV catalog, and the EU database EUVD (EUVD-2026-128497) lists no exploitation. DigitalOcean live-patched its own fleet and told customers: “No action needed on your part.”
Why it is dangerous
This is a use-after-free bug: software keeps using memory that has already been released. KVM confines each virtual machine with a table that translates the VM’s memory addresses into the server’s real memory. During a certain kind of reset of a nested VM (an INIT signal), KVM fails to restore its pointer to that table, so it keeps pointing at memory that has already been freed.
Think of a hotel guest handed a map showing only their own rooms. Because of the bug, the front desk is left holding a reference to a sheet that was thrown away, and the guest can draw their own map of the whole building on it. According to Terra Security’s analysis, a VM could reach any part of the host’s physical memory this way, which on a shared server includes other customers’ memory. On their own server the researchers also wrote 32 chosen bytes into the host kernel’s memory.
This cannot be exploited over the internet on its own. The attacker needs root inside a virtual machine on a vulnerable host, for example as a VPS customer or after taking over one of your VMs.
The Linux kernel’s CVE team scores it CVSS 8.8 out of 10 (CVSS 3.1), high severity with high impact on confidentiality, integrity and availability. Red Hat’s preliminary assessment gives 5.6 (Moderate) and treats the impact as denial of service only, Ubuntu rates it High, and DigitalOcean treated it as critical.
The CISA KEV catalog lists flaws with confirmed exploitation, and this one is not in it. The EPSS score estimates the chance of exploitation in the next 30 days and stood at just 0.19% on November 22, 2026.
Who is affected
Only servers acting as KVM hosts, with AMD processors and nested virtualization enabled, are affected. Nested virtualization means a VM can run VMs of its own. According to VirtFusion’s operator guide it is on by default in the kvm_amd module, so an unpatched host is exposed unless someone deliberately turned it off.
- The flawed code entered the 6.1 kernel in 2022 and was carried into the 5.15.81 and 6.0.11 releases, so older kernel lines and distribution kernels built on them are in range.
- Debian: kernels in all releases are vulnerable, with a fix only in unstable.
- Ubuntu: kernels listed as vulnerable, no fixed package yet.
- Red Hat Enterprise Linux 9 and 10: affected, fix deferred. RHEL 8 is not affected.
Distribution status is as of November 20, 2026, per The Hacker News. Typical environments are VPS and cloud providers, in-house KVM virtualization servers, and test setups that run virtualization inside VMs. Intel-based hosts are not affected. If you only rent a VPS, the fix is the provider’s job, and updating the system inside your VM changes nothing.
How to check if you are affected
Run these on the host, the physical server, not inside a VM. All of them only read information.
grep -m1 vendor_id /proc/cpuinfo
lsmod | grep kvm
cat /sys/module/kvm_amd/parameters/nested
uname -r
If vendor_id shows AuthenticAMD, the kvm_amd module is loaded and nested reads 1 or Y, the host is exposed until it runs a fixed kernel. If the nested file does not exist, kvm_amd is not loaded. Compare the uname -r output with the fixed versions, and for distribution kernels check your distribution’s security tracker. If you only rent a VPS, ask your provider.
How to fix it
The fix lives in the host kernel. Upstream, it shipped in stable kernel 7.3.5 and in the long-term kernels 6.18.64 and 6.12.118, as the Linux kernel CVE team’s announcement shows. Most distributions do not have fixed packages yet. The flaw is not in the CISA KEV catalog, so CISA has set no required action for it.
- First, list your KVM hosts with AMD processors and nested virtualization on. Prioritize those where other people have root inside VMs (customers, contractors, developers).
- If a fixed kernel is available, install it and reboot the host. Without a reboot the old kernel keeps running, so migrate VMs first or schedule downtime.
- If there is no fix yet, turn nested virtualization off. According to The Hacker News this is the only available protection and closes this path completely. Add this line to the host’s module configuration (for example a file in
/etc/modprobe.d/):
options kvm_amd nested=0
Then reboot the host, the setting cannot change while VMs are running. After the reboot, confirm with uname -r and cat /sys/module/kvm_amd/parameters/nested, which should show 0 or N. VMs will no longer be able to run VMs of their own, so warn anyone who relies on that beforehand.
Once a fixed distribution kernel is out, install it, reboot, and re-enable nested virtualization if you need it. Links to the actual kernel fixes are in the NVD record.
What if you may already be compromised
According to The Hacker News, there is no guidance for detecting a past attempt, and no one has reported exploitation outside research. The attack happens in memory and may not leave clear traces in logs. If a host was vulnerable and untrusted people had root in its VMs, check:
- unexpected crashes and kernel error messages on the host (
journalctl -k,dmesg) - unknown accounts, SSH keys and new services on the host
- unexpected kernel modules (
lsmod) - who had root inside VMs during that period and whether any of those VMs was compromised
If you have real grounds for suspicion, rotate passwords, keys and certificates of systems that ran on the host, since its memory may have held them. If you find anything suspicious or run shared infrastructure for customers, bring in a forensics specialist before reinstalling anything.