Patch available Security news
Linux KVM flaw CVE-2026-99312 lets a guest read host memory on AMD servers with nested virtualization
A bug in KVM, the hypervisor built into the Linux kernel, let a virtual machine on an AMD host with nested virtualization enabled reach the host's memory. Terra Security showed on 18 November 2026 that from a single DigitalOcean droplet it could map about 127 GB of host RAM. Fixed kernels shipped on 17 November; most distributions had no fixed package yet on 20 November. No exploitation in the wild is known.
Key points
- CVE-2026-99312: after an emulated INIT of a nested guest, KVM on AMD (SVM) did not reset the outer guest's MMU context, leaving it with a pointer to freed memory.
- Terra Security mapped about 127 GB of host RAM read-only from a DigitalOcean droplet and, on its own bare-metal server, wrote bytes into a live host kernel page.
- Fixed upstream on 2 Nov 2026 and shipped in kernels 7.3.5, 6.18.64 and 6.12.118 on 17 Nov. Debian had a fix only in unstable on 20 Nov.
- The kernel CVE team scores it CVSS 8.8; Red Hat's preliminary rating is Moderate (5.6). DigitalOcean says customers need not act.
Vulnerabilities: CVE-2026-99312
What happened
On 18 November 2026 Oren Yomtov of Terra Security published research into CVE-2026-99312, a flaw in KVM's x86 code that breaks the isolation between a virtual machine and its host on AMD servers with nested virtualization enabled. Terra reported it through DigitalOcean's bug bounty. On a DigitalOcean droplet the researchers mapped about 127 GB of host RAM read-only and found the host kernel's version banner in it. DigitalOcean says it confirmed the issue on its own hosts, treated it as critical and rolled out a live patch.
The Linux kernel CVE team published the CVE on 18 November. Fixed kernels had shipped the day before. No exploitation outside research is known.
Technical details
- Component: KVM,
arch/x86/kvm/x86.c, nested virtualization on AMD (SVM). - Bug type: when KVM emulates an INIT for a nested (L2) guest, it did not force an MMU reset, so the outer (L1) guest kept running with a stale MMU context that pointed to a page the host kernel had already freed. According to NVD, the fix forces an MMU reset on every emulated INIT.
- Origin: the regression came with the 2022 fix for CVE-2022-3344 and sat in the kernel for almost four years.
- Conditions: an AMD host with nested virtualization enabled and an attacker who controls a guest. Intel hosts and hosts without nested virtualization are not affected.
- Demonstrated impact: reading host memory (which on a shared host can include other customers' VMs) and, on the researchers' own bare-metal server, writing chosen bytes into a host kernel page. Terra does not claim code execution on the host.
Nested virtualization means running a hypervisor, and its own VMs, inside a virtual machine. Providers enable it for customers who need their own virtualization (Docker Desktop, Android emulators, test labs, WSL2). That is why shared hosts carry the most risk: a customer with full control of a guest is exactly the attacker this bug requires.
We do not describe how to exploit it.
Who is affected
- Hosts running KVM on AMD CPUs with
kvm_amdnested=1. - According to NVD, kernels since 6.1, plus the backports in 5.15.81+ and 6.0.11+. Fixed in 7.3.5, 6.18.64 and 6.12.118 (and 7.4-rc2). NVD lists no fix for the 5.15 and 6.1 long-term branches yet.
- Distributions (20 Nov 2026): Debian marks bookworm and trixie vulnerable, with a fix only in unstable (7.3.5-1). Red Hat says only nested virtualization on AMD KVM hosts is affected.
- Cloud customers generally need not act: the fix belongs on the provider's host. DigitalOcean says its customers need do nothing; Terra says many providers it surveyed were not exposed because they had not enabled nested virtualization on AMD.
How to tell if you were compromised
Neither Terra nor DigitalOcean has published indicators of compromise or detection guidance. Check exposure instead:
uname -r
cat /sys/module/kvm_amd/parameters/nested
A kernel older than the fixed versions together with 1 or Y from the second command means the host is exposed.
What to do
- Find AMD KVM hosts with nested virtualization enabled.
- Update the kernel to 7.3.5, 6.18.64, 6.12.118 or later, or install your distribution's fixed package once it ships, and reboot (or live-patch).
- Until then, disable nested virtualization: set
options kvm_amd nested=0in/etc/modprobe.d/and reload the module (stop the VMs first). The researchers and DigitalOcean name this as the workaround. - Prioritise multi-tenant hosts where untrusted customers control guests.
Compare the two CVSS vectors in our CVSS calculator: the kernel CVE team gives 8.8 (AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H), Red Hat a preliminary 5.6 (AV:L/AC:H/PR:L/UI:N/S:C/C:N/I:N/A:H).
Timeline
FAQ
I rent a VPS. Do I need to do anything?
Usually not - the fix belongs on the provider's host, not in your guest. DigitalOcean says it live-patched its fleet. If your provider runs AMD hosts with nested virtualization, ask whether it has patched.
How do I know whether nested virtualization is on?
On an AMD host, cat /sys/module/kvm_amd/parameters/nested prints 1 (or Y) when it is enabled. If it prints 0 or N, this path is closed.
Why do the CVSS scores differ so much?
The kernel CVE team rates high impact on confidentiality, integrity and availability (8.8). Red Hat's preliminary assessment counts only denial of service (5.6). The researchers did demonstrate reading host memory and, on their own server, writing to it.
Sources
Spotted a mistake? Let us know and we will correct it openly.