---
title: A guest-to-host hypervisor escape is disclosed (e.g. KVM CVE-2026-99312): find exposed hosts, cut the vulnerable feature, patch and reboot, and look for misuse | ctxsec
url: https://ctxsec.com/playbooks/hypervisor-guest-to-host-escape-disclosed-kvm-fleet-response
---

[← Playbooks](/playbooks)

# A guest-to-host hypervisor escape is disclosed (e.g. KVM CVE-2026-99312): find exposed hosts, cut the vulnerable feature, patch and reboot, and look for misuse

For teams running their own KVM/Proxmox/OpenStack hosts or offering VMs to others. Built from the Nov 2026 KVM AMD nested-virtualization escape, where a root user inside one VM could read about 127 GB of host memory.

For: Virtualization / platform engineers, Hosting & MSP operators, SOC

[1. Scope which hosts are actually exposed](#s1)[2. Cut the attack path before the patch](#s2)[3. Patch and actually reboot](#s3)[4. Look for signs it was used](#s4)

1.

Step 1 of 4

__

### Scope which hosts are actually exposed

Escapes usually depend on CPU vendor, a feature flag and a kernel range.

   - List hypervisor hosts with CPU vendor, kernel version and whether the vulnerable feature is on (for CVE-2026-99312: AMD + /sys/module/kvm_amd/parameters/nested = 1/Y)
   - Rank by who controls the guests: multi-tenant and customer VMs first, then CI runners and sandboxes that run untrusted code, then internal-only VMs
   - Check your distro's CVE tracker for a fixed kernel; upstream fixes often land before distro packages

2.

Step 2 of 4

__

### Cut the attack path before the patch

Disabling the feature is often cheaper than an emergency kernel rollout.

   - If tenants don't need nested virtualization, set options kvm_amd nested=0 and plan reboots (the parameter can't change with VMs running)
   - Tell tenants who rely on nested VMs before the change; move them to patched hosts first
   - Pause onboarding of new untrusted guests onto unpatched hosts

3.

Step 3 of 4

__

### Patch and actually reboot

A kernel on disk protects nothing until the host boots it.

   - Live-migrate guests off, install the fixed kernel (or a vendor live patch that names the CVE), reboot
   - Verify with uname -r and by re-reading the feature flag; record each host as done
   - Keep a list of hosts that can't reboot yet, with an owner and date

4.

Step 4 of 4

__

### Look for signs it was used

There may be no specific detection, so look for the preconditions.

   - Review which guests on exposed hosts had root and ran their own nested hypervisor (e.g. /dev/kvm used inside the guest)
   - Check host logs for unexplained KVM warnings, MMU errors or host kernel crashes around nested guests
   - If a hostile tenant is plausible, treat co-resident guests' secrets as exposed and rotate them

Original content by ctxsec. Printable checklist: use your browser's print view.
