The Short Version
What it is: CVE-2026-64561, named Zapscape by its finder, Hyunwoo Kim (@v4bel), is a use-after-free in the Linux KVM x86 shadow memory management unit. Two page fault paths validate the shadow root before reclaiming shadow pages instead of after, so KVM keeps building mappings beneath a root reclamation has already invalidated. The result is a full guest-to-host escape. The published PoC documents a path to running code as root on the host, not just a panic that kills every co-tenant VM. Jump to this section →
Potential impact: The researcher reports a path from an attacker-controlled guest to host kernel code execution, which would be a full escape from a virtual machine onto its host. The same disclosure describes two further outcomes, a host panic that interrupts every workload running on the affected machine, and a separate local privilege-escalation path that starts from an ordinary account on the host where /dev/kvm is world-writable. Jump to this section →
Severity: Red Hat rates it Important with a CVSS 3.1 base score of 7.0 under CWE-825. Exploitation needs guest kernel privileges, nested virtualization, and a compatible host configuration. Red Hat’s own wording frames the impact as instability and denial of service in virtualized environments, which is a materially softer read than the researcher’s. Jump to this section →
Who is exposed: Hosts running KVM with nested virtualization enabled, on a kernel that contains the change introduced by commit f95eec9bed76 in Linux 5.9 and does not yet contain upstream fix 2abd5287f083. AMD SVM systems using nested page tables are the most directly reachable. Intel systems require both four-level and five-level EPT page walks to be exposed to the first-level guest, which limits that path to Ice Lake-SP and newer. Jump to this section →
What has been verified: The public proof of concept has not produced a guest-to-host escape on CloudLinux-built kernels in our testing. It does not run against them without modification, and in-house reproduction has so far yielded only a crash of the first-level guest. Treat the escape as credible, not as a demonstrated fleet-wide compromise. Jump to this section →
What to do now: Start by finding out whether a host is in scope at all.
cat /sys/module/kvm_amd/parameters/nested 2>/dev/null
cat /sys/module/kvm_intel/parameters/nested 2>/dev/null
A 1 or Y means the loaded KVM module permits nested virtualization, which puts the host inside the potential attack surface. It does not prove exploitability on its own, since kernel patch state, host CPU capabilities, the virtual CPU model exposed to L1, and whether an untrusted guest can create L2 all still matter.
No output usually means the matching KVM module is not loaded right now, which is not the same as out of scope, so check how the host is actually used and whether KVM can be loaded later. Read the file rather than assuming nobody turned nesting on, because upstream KVM has enabled the nested parameter by default for both Intel and AMD since Linux 4.20, and a distribution can override that default in either direction.
On hosts that do not need nested virtualization, disable it, since that removes the prerequisite the entire chain depends on. Check /dev/kvm permissions as separate hardening while you are in there.
KernelCare closes this flaw in the running kernel with no reboot and no VM restart. The first rebootless patches have now shipped: Debian 13 is in the main feed, and Ubuntu 24.04, Ubuntu 24.04 for AWS, Ubuntu 22.04 for AWS 6.8 kernels, and Proxmox VE 8 have builds in the test feed expected to promote to main. The remaining families are still in build and review, with availability landing per kernel and feed as validation completes. Apply a fixed vendor kernel where one already exists for your platform.
Close Zapscape Without an Emergency Host Drain
KernelCare applies the upstream MMU reordering fix directly to the running kernel, so the vulnerable path closes without stopping tenant workloads.
Status as of August 11, 2026: Debian 13 is covered in the main feed, with Ubuntu 24.04, Ubuntu 24.04 for AWS, Ubuntu 22.04 for AWS 6.8 kernels, and Proxmox VE 8 in the test feed and expected to promote to main. Every Enterprise Linux 9 kernel is covered by patches in review apart from the oldest 70-series, which has been deferred, Debian 12 is in review awaiting hardware validation, Enterprise Linux 8 and Ubuntu 22.04 are prepared with validation pending, and Enterprise Linux 10 is queued. Coverage has to exist for your exact kernel, so confirm it before dropping a temporary mitigation.
The Long Version
Two descriptions of Zapscape are in circulation, and they are very far apart. The researcher describes a chain that ends in host kernel code execution from inside a guest. Red Hat describes instability in virtualized environments. Both are talking about the same commit.
The honest answer sits between them, and the detail below explains why. What has been reproduced in house is narrower than the disclosure suggests. The public proof of concept can crash the attacker-controlled L1 guest in some environments, while the researcher documents a configuration-dependent path onward to host code execution. That gap is worth stating plainly, and neither reading is an argument for leaving a broken virtualization boundary unpatched.
Why Zapscape Matters
Nested virtualization lets a virtual machine run virtual machines of its own. A cloud customer might use it to run a hypervisor, test infrastructure software, operate a CI environment, or enable Windows features that depend on Hyper-V.
That produces a three-level arrangement:
- L0 is the physical host
- L1 is the attacker-controlled guest
- L2 is a virtual machine created inside L1
Zapscape abuses memory management that L0 performs while servicing activity from L2. The attacker starts with root inside L1 and creates the nested guest needed to drive the vulnerable KVM path.
Root inside a cloud virtual machine is not unusual. Customers normally control their own guest operating systems, and the boundary is supposed to stop that access from reaching the hypervisor. Zapscape targets the boundary itself.
How CVE-2026-64561 Works
KVM uses shadow page tables to translate guest memory accesses into host mappings, and reclaims those structures once it hits its quota for active MMU pages.
The vulnerable fault paths, direct_page_fault() and FNAME(page_fault)(), called is_page_fault_stale() before make_mmu_pages_available().
That order is the bug.
Reclamation inside make_mmu_pages_available() can zap the root KVM is currently using and mark it invalid. Because the stale-root check already ran, KVM carries on building mappings beneath a root that is no longer valid. Child shadow pages inherit their parent’s role, so they inherit the invalid state, and invalid pages end up back on the list of active MMU pages.
That is the invariant commit f95eec9bed76 added in Linux 5.9 to prevent. Once it breaks, a later recursive cleanup through mmu_page_zap_pte() runs with no root_count guard and operates on structures that have already been freed.
The proof of concept drives the sequence by switching memory modes inside L1 to alias shadow pages while memory activity from L2 pushes L0 into quota-based reclaim. The researcher then turns the resulting corruption into host kernel code execution.
The ordering problem itself dates back to 2008. It became security-relevant in Linux 5.9, when the invariant it violates arrived. Nobody wrote a bug in 5.9. The kernel simply started depending on something that had never been true.
Is the KVM Escape Confirmed?
The researcher supplies a guest-side proof of concept and documents a complete path to host code execution. Exploitability, though, is heavily configuration-dependent. It targets AMD and was developed against a specific Linux 7.1.3 configuration under QEMU TCG rather than on production hardware, so reaching a real host would mean rewriting the guest-side actions and adapting to the target kernel configuration.
Our testing has not reproduced the full escape on any CloudLinux-built kernel. The public exploit does not run against them without modification, and the observed result is a crash of the L1 guest, which is a denial of service for the attacker’s own virtual machine rather than host compromise.
Enterprise Linux 8 deserves a note for anyone testing this themselves. The public proof of concept does nothing at all there, not even a crash, so it cannot serve as a before-and-after oracle on that family. Validate EL8 kernels against a general KVM test suite instead of concluding from silence.
None of that proves the affected kernels are safe. It means the vulnerability should be described precisely. Zapscape presents a credible guest-to-host escape path, and the public exploit is not currently a universal escape for every kernel that contains the vulnerable code.
Who Is Affected
The vulnerable code path exists from the Linux 5.9 change in commit f95eec9bed76, dated July 8, 2020, through kernels that have not received fix 2abd5287f083, which landed upstream on July 21, 2026.
Reachability additionally requires:
- KVM to be in use
- Nested virtualization to be enabled
- Kernel-level control inside the L1 guest
- Host CPU and virtual CPU features compatible with the attack
On AMD hosts using SVM and nested page tables, the disclosure identifies no additional CPU-feature prerequisite beyond the main requirements already listed. Intel systems are reachable only where both four-level and five-level EPT page-walk lengths are exposed to L1, which generally means Ice Lake-SP and newer with those capabilities passed through.
Affected or potentially affected platform families identified during our analysis:
- Enterprise Linux 10, AlmaLinux 10, and CloudLinux 10, where the CentOS Stream 10 backport is still in draft
- Older Enterprise Linux 8 and 9 kernels that do not contain the vendor fix, including EL8-derived flavors
- Debian 11, 12, and 13
- Ubuntu 22.04 and 24.04
- Oracle UEK7
- Proxmox VE 7 and VE 8
Already corrected or out of scope:
- RHEL 8 kernels from
4.18.0-553.147.1.el8_10onward, shipped July 23, 2026 - RHEL 9 kernels from
5.14.0-687.30.1.el9_8onward, shipped July 22, 2026 - RHEL 6 and RHEL 7, including kernel-rt, which predate the Linux 5.9 change that made the flaw security-relevant
- Any host not running KVM, or running it with nested virtualization disabled
The RHEL entries come with a catch. Red Hat shipped the correction quietly, batched with other vulnerability work, and its public CVE page continued to list RHEL 8 and RHEL 9 as affected after those errata went out. Changelog evidence is the stronger signal, and page state has been known to flip within hours, so re-check Red Hat’s current data before quoting an affected state to a customer or an auditor.
The same caution applies more generally. Vendor kernels backport individual changes constantly, so an upstream version number is not an answer about exposure. Check the advisory and patch status for the exact running kernel.
How to Mitigate Zapscape
Every prerequisite for this bug is something that can be taken away. All three of these are temporary controls, not substitutes for the fix.
- Disable nested virtualization. The most direct option, because the chain requires L1 to create and operate a virtual machine of its own.
# /etc/modprobe.d/99-zapscape.conf
options kvm_amd nested=0
options kvm_intel nested=0
The module parameter is read at load time, so applying it means stopping every running virtual machine and reloading kvm_amd or kvm_intel. On a production host, that is a drain or a reboot rather than a live change.
If the fleet runs ordinary single-level virtual machines, apply it and move on, since nothing in that workload needs nested virtualization. If the host deliberately provides it, think nested hypervisor tenants, CI systems that boot virtual machines, virtualized development and test environments, or Windows guests using VBS or Hyper-V, leave it enabled and prioritize patching instead. Assess workload impact before touching a host you did not build.
- Restrict access to
/dev/kvm. This does nothing about an attack from an existing guest. It closes the separate path where an ordinary local user on the host reaches KVM directly.
ls -l /dev/kvm
If the device shows mode 0666, tighten it and grant legitimate users access through the kvm group. Make that change in a udev rule rather than with a bare chmod, since udev resets the mode on the next boot or device event.
- Review five-level EPT exposure on Intel. Only relevant where both EPT capabilities reach L1.
grep -w ept_5level /proc/cpuinfo
No output means the host lacks the five-level EPT capability the disclosed Intel path needs. If the flag is present, the next question is whether both four-level and five-level EPT are actually reaching the untrusted L1 guest, since /proc/cpuinfo describes the host CPU rather than what the virtual CPU model passes through. Where nested virtualization must stay on, avoid CPU models that expose five-level EPT to untrusted guests. On newer Intel systems, disabling nested virtualization is still the more dependable temporary control.
Once protection is in place, remove the temporary config with sudo rm /etc/modprobe.d/99-zapscape.conf and reload the module during the next maintenance window.
The Fix
The upstream correction is commit 2abd5287f083, “KVM: x86: Check for invalid/obsolete root after making MMU pages available.”
The change is small. It moves the stale-root check to after make_mmu_pages_available() in both affected fault paths, under a dozen changed lines across arch/x86/kvm/mmu/mmu.c and arch/x86/kvm/mmu/paging_tmpl.h. KVM now performs any reclamation first and then confirms the root is still valid before continuing, which stops the hypervisor from creating active mappings beneath a root that reclamation has already invalidated.
No new data structure, no persistent state to track. That is what makes it suitable for rebootless patching, though the backport itself varies with how far each vendor kernel’s fault paths have drifted from mainline.
KernelCare Patch Status and Deployment
As of August 11, 2026, the first releases are out:
Debian 13: main feed
Ubuntu 24.04 Noble, Ubuntu 24.04 Noble for AWS, Ubuntu 22.04 Jammy for AWS, 6.8 kernels, and Proxmox VE 8 are in the test feed
The test-feed builds were expected to reach the main feed the same day, so check the feed state for your kernel rather than assuming these are still test-only.
Remaining family status:
Every Enterprise Linux 9 kernel is covered by patches in review, apart from the oldest 70-series kernels, which need adaptation and have been deferred to separate work
- Ubuntu 24.04 patches are ready and in review
- Debian 12 patches are in review and awaiting hardware validation
- Enterprise Linux 8 and Ubuntu 22.04 patches have been prepared, with validation still pending
- Enterprise Linux 10 is in the queue
AMD test hardware, not patch complexity, has been the limiting factor across several families. Availability will differ by kernel until testing and the completion of the roll-out.
If you are eager to update while a patch is still in the testing feed:
kcarectl --update --prefix test
After roll-out to the production feed, the patch arrives on the next scheduled run, or on demand:
kcarectl --update
Then check what is actually loaded on the running kernel:
kcarectl --patch-info | grep 'CVE-2026-64561'
A build timestamp is the secondary check:
kcarectl --info | grep kpatch-build-time
A build dated August 6, 2026 or later is a useful signal, not a guarantee that this specific CVE is covered for that kernel. Where the patch information does not yet list the CVE, confirm coverage for the exact kernel through customer support rather than reading the date as proof. Either way, verify before removing a temporary mitigation, not after.
Learn how KernelCare patches critical kernel vulnerabilities without rebooting →
Timeline
| Date | Event |
| 2008 | The stale-root check is ordered before shadow page reclamation, harmless at the time |
| July 8, 2020 | Commit f95eec9bed76 enters Linux 5.9 and adds the invariant that the ordering violates |
| July 21, 2026 | Upstream fix 2abd5287f083 lands |
| July 22 and 23, 2026 | Red Hat ships the correction in RHEL 9 and RHEL 8 errata, batched with other vulnerability work |
| August 6, 2026 | CVE-2026-64561, the technical disclosure, and the guest-side proof of concept become public |
| August 6, 2026 onward | Vendor updates and KernelCare live patch validation continue across affected kernel families |
| August 11, 2026 | First KernelCare rebootless patches release, Debian 13 to the main feed and Ubuntu 24.04, Ubuntu 24.04 for AWS, Ubuntu 22.04 for AWS 6.8 kernels, and Proxmox VE 8 to the test feed |
References
- Hyunwoo Kim (@v4bel), Zapscape disclosure to oss-security
- Zapscape technical write-up and proof of concept, zapscape.io
- Upstream Linux fix
2abd5287f083, “KVM: x86: Check for invalid/obsolete root after making MMU pages available” - Linux commit
f95eec9bed76, “KVM: x86/mmu: Don’t put invalid SPs back on the list of active pages” - Red Hat advisory and CVE record for CVE-2026-64561
- CentOS Stream 10 backport merge request 3068, in draft
- Applicable RHEL, AlmaLinux, CloudLinux, Debian, Ubuntu, Oracle, and Proxmox advisories