Linux KVM CVE-2026-99312: On AMD Hosts With Nested Virtualization On, a Guest Can Map Host Memory. Fixed in 7.3.5, 6.18.64 and 6.12.118, or Set nested=0 Today
What to do first. If you run Linux KVM hosts on AMD processors and let tenants or untrusted users run VMs on them, check two things today: your host kernel version and whether the kvm_amd module has nested virtualization turned on. Fixed kernels are stable 7.3.5, 6.18.64 and 6.12.118 (released Tuesday, November 17, 2026). If you cannot reboot into one yet, turn nested virtualization off (set the kvm_amd parameter nested to 0 and reboot the host). The researchers say that "closes this path completely." Patching guest operating systems does nothing: this is a host kernel bug.
Intel hosts are not affected. Hosts that never enabled nested virtualization are not exposed to this path, but note that kvm_amd ships with nested turned on by default, so "we never touched it" usually means "it is on."
What the bug is
The kernel's own description, published with the CVE on Wednesday, November 18, is titled "KVM: x86: Unconditionally reset MMU context when emulating INIT." When a guest that is itself running a nested guest (KVM calls them L1 and L2) receives an INIT reset while the nested guest runs with paging off, KVM skips resetting its memory-management context and keeps running L1 with a stale one.
Why AMD only, in the kernel's words: "this bug only affects nested SVM as INIT is an unconditional intercept on VMX, i.e. it's impossible to emulate INIT while L2 is active." SVM is AMD's virtualization extension; VMX is Intel's.
Terra Security researcher Oren Yomtov found it and published the analysis on November 19. Per Terra, the stale pointer is the outer VM's nested page table root, which names a page KVM has already freed. KVM later reuses that page as a page table the guest can influence, and the guest then forges entries that map host physical memory. In Terra's words: "Our VM stays confined only as long as the NPT pointer names a page KVM still owns."
The root cause is old. The kernel CVE record names the introducing commit as one that landed in 6.1 in late 2022 (the fix for CVE-2022-3344) and was backported into 5.15.81 and 6.0.11. DigitalOcean says it sat in the kernel "for just under four years."
What an attacker needs
A VM on a vulnerable AMD KVM host with nested virtualization enabled.
Kernel-level control of that VM. The Hacker News reports both of Terra's tests started from root in a researcher-controlled VM; running a nested guest and delivering INIT requires control of the guest kernel. The kernel CNA's CVSS vector is local with low privileges (AV:L/PR:L), scope changed.
Network access alone does not get anyone in. The risk is to whoever lets strangers rent or run VMs: VPS and cloud providers, CI runners that hand out VMs, hosting panels, labs.
What Terra demonstrated: on a DigitalOcean Droplet, mapping about 127 GB of host RAM read-only; on their own bare-metal AMD server, writing 32 bytes into a live host kernel page, confirmed by a witness module. A host memory write is the basis of a full guest-to-host escape.
Severity: two scores, read both
Linux kernel CNA: CVSS 3.1 8.8, high confidentiality, integrity and availability (the score on the NVD record, which NVD lists as "Received," not yet analyzed).
Red Hat: 5.6, Moderate, preliminary, availability only, because "successful exploitation requires nested virtualization."
Terra's demonstrated host memory read and write is more than availability. If your hosts run AMD with nested on, plan from the 8.8.
Exploitation and proof of concept
No public exploit code: Terra described the technique and its two demonstrations but did not publish code. We found no report of use outside research; The Hacker News cites VirtFusion saying the same. CISA's KEV catalog does not list it (checked November 20). That can change quickly now that the mechanism is public, so treat "no PoC" as a short window, not a reason to wait.
Which kernels are fixed
From the kernel CVE announcement and NVD record:
Fixed: 7.3.5, 6.18.64, 6.12.118, and mainline 7.4-rc2 (fix commit c7c288d178cb).
Affected: 6.1 and every later line until those fixes; 5.15.81 and later in the 5.15 line; 6.0.11 and later in 6.0.
No fixed release is listed yet for the 5.15, 6.1 or 6.6 lines. If your distribution kernel is based on one of those, wait for your vendor's backport or use the nested=0 mitigation.
Distribution status, as read on November 20:
Debian security tracker: bookworm (6.1.192-2), trixie (6.12.117-1) and forky (7.3.4-1) listed vulnerable; only sid (7.3.5-1) fixed.
Red Hat: Moderate, per its CVE page; The Hacker News reports RHEL 9 and 10 affected with the fix deferred and RHEL 8 not affected (we did not see that product table on the Red Hat page itself).
Ubuntu: The Hacker News reports it rated High with no fixed package yet. We did not verify this on Ubuntu's tracker.
Who has said they are patched
DigitalOcean: was vulnerable (Terra demonstrated the read on a Droplet and DigitalOcean validated it on its own hosts), live-patched its fleet, and tells Droplet customers "No action needed on your part."
Terra says "many cloud providers we surveyed were not vulnerable because they did not enable nested virtualization on AMD," but names none of them. We found no statement from AWS, Microsoft Azure or Google Cloud and make no claim about them either way.
VirtFusion, a hosting control panel, published an operator guide for its customers.
Am I affected?
On the host (as root):
Kernel version: run uname -r and compare against the fixed list above, or your distribution's tracker.
CPU vendor: run grep -m1 vendor_id /proc/cpuinfo. AuthenticAMD means this bug applies; GenuineIntel means it does not.
Nested setting: run cat /sys/module/kvm_amd/parameters/nested. 1 or Y means nested virtualization is on and the host is exposed until patched. 0 or N means this path is closed. If the file does not exist, kvm_amd is not loaded and this host is not running AMD KVM guests.
Inside a guest you rent, you cannot fix this yourself; it is your provider's host. Running grep -c svm /proc/cpuinfo in the guest shows whether the host passes AMD virtualization through to you (a nonzero count means nested virtualization is exposed to your VM). Ask your provider whether its hosts carry the fix.
Mitigation if you cannot patch yet
Add the line options kvm_amd nested=0 to a file in /etc/modprobe.d/ (for example /etc/modprobe.d/kvm-amd.conf), then reboot the host and confirm the nested parameter reads 0 or N. Warn tenants first: any tenant running its own hypervisor or hardware-accelerated emulator inside your VMs loses that ability.
What we checked on our side
No edge check: this is a host kernel flaw reached from inside a VM, not something anyone can send over HTTP, so our web edge cannot see it. No indicators: the primaries publish no IPs, domains or hashes, and we added nothing to the feed.
How we know
Kernel CVE announcement (linux-cve-announce, November 18) and the NVD record for CVE-2026-99312 for the description, affected ranges, fix commits and CNA score; Terra Security's write-up for the mechanism, demonstrations, mitigation and timeline (reported to KVM maintainers October 21, fix merged November 2, stable releases November 17); DigitalOcean's post for its exposure and patch; the Debian security tracker and Red Hat CVE page for distribution status; The Hacker News for Ubuntu, RHEL product detail and the VirtFusion exploitation note. We found no oss-security post for this CVE.
Sources
Terra Security: https://www.terra.security/blog/kvm-guest-to-host-escape
Kernel CVE announcement: https://lists.openwall.net/linux-cve-announce/2026/11/18/2
NVD: https://nvd.nist.gov/vuln/detail/CVE-2026-99312
Fix commit (mainline): https://git.kernel.org/stable/c/c7c288d178cb0b78d3464a616b16b1aa0b316049
DigitalOcean: https://www.digitalocean.com/blog/cve-2026-99312-kvm-patch
Red Hat: https://access.redhat.com/security/cve/cve-2026-99312
Debian: https://security-tracker.debian.org/tracker/CVE-2026-99312
VirtFusion: https://www.docs.virtfusion.com/6.2/security/cve-2026-99312/
The Hacker News: https://thehackernews.com/2026/11/new-linux-kvm-flaw-cve-2026-99312-lets.html
Read next: ITScape: the KVM arm64 guest escape that reaches the host kernel, the same class of bug on the other architecture.
Every indicator in this post is in the feed. Free.
1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.



Comments