CVE-2026-99312: A KVM Nested-AMD Bug Scores 8.8, With Fixes Listed for Three Kernel Branches
Security / news
CVE-2026-99312: A KVM Nested-AMD Bug Scores 8.8, With Fixes Listed for Three Kernel Branches
The Linux CNA's scoring notes describe a freed page-table root that a guest can aim at; the commit message describes only a stale MMU context.

A local user who controls a guest virtual machine on an AMD host (nested virtualization on, which is the kvm_amd default) can trigger a stale-page-table condition in the Linux kernel's KVM code, according to the scoring notes in CVE-2026-99312. The Linux kernel CNA rates it 8.8 (CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H). Intel hosts are not affected.
The mitigation is a kernel update. The CVE record lists fixed versions 6.12.118, 6.18.64 and 7.3.5, plus 7.4-rc2 on mainline.
What the CVE-2026-99312 record says is broken
The patch is titled "KVM: x86: Unconditionally reset MMU context when emulating INIT". Per its commit message, KVM skipped an MMU reset on an INIT event when paging was already disabled. If the vCPU was running a nested guest (L2) at that moment, KVM went on running the outer guest (L1) with a stale MMU context.
The commit message says the bug "only affects nested SVM", AMD's virtualization extension. On Intel's VMX, INIT is an unconditional intercept, so it is impossible to emulate INIT while L2 is active.
Affected and fixed kernels
The record marks the flaw as introduced by commit ed129ec9057f ("KVM: x86: forcibly leave nested mode on vCPU reset"). It lists the following.
| Kernel line | Status in the CVE record | Fixed in |
|---|---|---|
| 5.15 (from 5.15.81) | Affected | None listed |
| 6.0 (from 6.0.11) | Affected | None listed |
| 6.1 onward | Affected | See below |
| 6.12 | Affected | 6.12.118 |
| 6.18 | Affected | 6.18.64 |
| 7.3 | Affected | 7.3.5 |
| Mainline | Fixed | 7.4-rc2 |
The record lists no fixed release for the 5.15 line or for 6.1 itself. Operators on those branches should check their distribution's advisory rather than assume a backport exists.
The score overstates the reach and may understate the impact
The vector is local, needs low privileges and no user interaction. It is not network reachable, and the CNA's own scenario text says no network message carries the trigger. The attacker needs a non-root user able to create a VM and run a nested hypervisor inside it.
Where that precondition holds, the CNA's scenario text describes more than a crash. It says a freed page-table root stays referenced by the VMCB's nested_cr3 field, so the CPU walks a freed host page, which could disclose host page-frame numbers or let the guest write arbitrary host physical pages. It also describes a livelock of the vCPU thread until the freed page is reused.
The commit message itself describes only the stale MMU context. The memory-corruption claims come from the scoring rationale, and no public proof of concept turned up in the sources consulted. No source fetched for this report says the flaw is exploited in the wild.
Timeline from patch to CVE
- Sept. 25, 2026: the CVE ID is reserved by the Linux CNA.
- Oct. 21, 2026: Sean Christopherson of Google authors the fix. The commit credits reports from Stefan Teodorescu, Oren Yomtov and Chris Mason, and carries a stable-kernel tag.
- Nov. 18, 2026: the CVE record is published. NIST's National Vulnerability Database lists the entry with status "Received", meaning NIST has not yet run its own analysis. The 8.8 score is the CNA's.
- Nov. 19, 2026: the record is updated again.
What operators should do
Check whether hosts run AMD processors with kvm_amd loaded and nested virtualization enabled. Where guests do not need to run their own hypervisors, turning nested virtualization off removes the precondition, though the record does not name that as an official workaround.
Everyone else should move to 6.12.118, 6.18.64 or 7.3.5, or the equivalent distribution build. Multi-tenant hosts that let customers run nested guests should do this first. For the 5.15 and 6.1 lines, there is nothing in the record to install yet.
Related patching decisions on this site: nginx2root shows a similar gap between a fix landing and distributions shipping it, and the Java4Shell coverage shows what happens when vendors disagree about fixed builds.
Sources
More in Security
- 01Ruby2Shell (CVE-2026-132825): Ruby Is Fixed, but GitLab, GitHub Enterprise and Chatwoot Are Listed Without a PatchA one-byte heap overflow in Ruby's case conversion carries a 7.4 local score, while the researchers say stock Rails is remotely reachable and will publish a proof of concept Nov. 25.
- 02NPCI Wants Overlay Permissions Restricted on Sideloaded Apps; Google's 2024 India Block List Omits ThemThe reported ask covers accessibility, device admin and display-over-apps. The pilot Google announced on Oct. 3, 2024 named four other permissions.
- 03CISA's Nov. 18 KEV Additions Include Apache Axis 1.x, Which Has No Fix, With a Nov. 21 DeadlineThree of four new entries carry a three-day federal deadline; LiteLLM's advisory names two different patched versions.
- 04NSANet Went Down Nov. 17 With 'Routine Maintenance' as the Only Cause on RecordOfficials from the White House to the Senate Intelligence Committee say they see no attack, but the NSA has not said which systems failed or for how long.