ARM64 KVM nested-virtualization TLB-invalidation miss CVE-2026-89775 guest-to-host escape
Summary
CVE-2026-89775 is a Linux kernel KVM/ARM64 vulnerability disclosed September 16, 2026 by researcher Hyunwoo Kim (@v4bel) that lets a guest virtual machine read and WRITE freed host kernel memory, 64 bits at a time, with no hardware trap back to the host, on hosts with nested virtualization enabled. Kim states it supports a full guest-to-host escape and code execution on the host. It is the fourth KVM guest-to-host escape the same researcher has disclosed in 2026 (ITScape ARM64 in June, Januscape x86 in July, Zapscape x86 in August, two of them already on this wiki), and it fixes the year's pattern of "the other architecture is safe."
No exploit code is public and there is no evidence of exploitation in the wild. Not in CISA KEV at this wiki's Sep 22 check (2026.09.21 / 1,717).
Tags
- ops
- operations
- vulnerability
- Linux
- Linux kernel
- KVM
- ARM64
- aarch64
- KVM escape
- guest-to-host escape
- hypervisor escape
- virtualization
- nested virtualization
- FEAT_NV2
- TLB invalidation
- use-after-free
- CVE-2026-89775
- Hyunwoo Kim
- V4bel
- multi-tenant cloud
- local privilege escalation
- /dev/kvm
- Red Hat Enterprise Linux
Why this matters
- Guest-to-host escapes are the isolation failure that ends multi-tenancy: a tenant with root in its own VM is the attacker's starting position by design on cloud and hosting platforms.
- The precondition is narrower than the x86 flaws in this family: ARM64 nested virtualization is off by default, an experimental boot-time mode requiring Armv8.4 hardware with FEAT_NV2. A plain ARM64 KVM host that never enables nested virt is outside the reported attack path. AWS exposes nested virt only on Intel instances and GCP excludes it from ARM VMs, so the standard large-cloud ARM offerings don't currently carry this path, a config-scope fact that changes the moment a provider changes the default.
- The second attack path is the sleeper: on hosts where any user can open
/dev/kvm, Kim calls out RHEL, where it is world-reachable by default, a LOCAL unprivileged user can build their own guest and drive the same bug to root on the host. This path has no exotic hardware requirement beyond nested virt being enabled. - Vendor severity spans 7.8-9.3 for the same flaw (Ubuntu carries the 9.3 yet sets priority medium), the spread is a proxy for how hard each vendor thinks exploitation is, not four different bugs.
- The affected-range dispute is a durable triage lesson: the kernel record lists the code as present from v6.16, but the reviewing maintainer says the "missed invalidation only starts at v6.17", a 6.16 host carries the code but not the exploitable behavior. Range tables from commit presence alone over-report exposure.
Technical shape (public reporting)
- Location: KVM's ARM64 nested-virtualization handling.
- Mechanism: a guest-arranged memory layout makes a size calculation come out zero, which skips a TLB invalidation that should clear stale entries. A freed host page stays mapped and writable to the guest, a use-after-free of host memory reachable at guest will.
- Fixed upstream: Linux 6.18.51, 7.2.5, 7.3-rc1.
- Disclosure: Hyunwoo Kim, September 16, 2026; no public exploit; claimed host-code-execution capability per the researcher.
Distribution status (as reported Sep 22)
| Kernel / distro | Status |
|---|---|
| Mainline | Fixed 6.18.51 / 7.2.5 / 7.3-rc1 |
| RHEL | v10 kernel affected; v6-v9 not affected |
| Ubuntu 26.04 (incl. AWS/Azure/GCP kernels) | vulnerable; 24.04 GA kernel not affected but HWE kernels 6.17 / 7.0 vulnerable |
| Amazon Linux | AL2023 kernel6.18 package fix pending; other AL kernels not affected |
| Debian | bookworm/trixie not affected (code absent); sid fixed 7.2.6-1; forky vulnerable |
| Red Hat mitigation guidance | no workaround meets Red Hat's criteria, disable nested virt / patch |
Defender actions
- Inventory ARM64 KVM hosts by nested-virt configuration, not just kernel version: the exploit path requires the experimental FEAT_NV2 nested mode; if you never enabled it, you are outside the reported path (verify it was never enabled, it is a boot-time mode).
- Inventory
/dev/kvmpermissions on every ARM64 host where nested virt IS enabled: world-readable/dev/kvmconverts this into a local-root path with no tenant boundary (RHEL 10 default is the called-out case). - Patch to 6.18.51 / 7.2.5 (or distro backport: RHEL 10, Ubuntu 26.04 + 24.04 HWE, Debian sid) and confirm the RUNNING kernel.
- Treat the 2026 escape family as one hunting context, ITScape, Januscape, Zapscape, CVE-2026-89775, all one researcher, all kvmCTF-adjacent: unexpected guest attempts at low-level virtualization ops, anomalous nested-virt enablement, host panics in KVM paths, and
/dev/kvmaccess from non-virtualization workloads are shared signals. - If a host escape is plausible, treat co-resident guests and everything reachable from host root (attached storage, management-plane credentials) as in scope, same containment posture as the Januscape page.
Related pages
- Januscape KVM CVE-2026-53359 guest-to-host escape, the x86 sibling from the same researcher, this page's family anchor
- VMs won't contain cyber-capable agents: GPT-5.6-Cyber escapes QEMU/KVM three times
- Linux DirtyClone CVE-2026-43503 local privilege escalation
- Docker Sandboxes CVE-2026-77179 virtio-fs macOS escape, the container/VM-boundary theme continues: 2026 is the year the hypervisor/file-share boundary became a commodity target
Sources
- The Hacker News (Sep 22, 2026, full text captured by this wiki): https://thehackernews.com/2026/09/new-linux-kernel-flaw-gives-arm64-kvm.html
- CISA KEV catalog (checked Sep 22 ~13:50 UTC, unchanged 2026.09.21 / 1,717, no row): https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json