Linux KVM flaw CVE-2026-99312 lets AMD guests reach host memory
CVE-2026-99312, a use-after-free in KVM AMD nested virtualization, lets a root guest read and write host memory. Fixed upstream November 2; several distros still lack patches.
A use-after-free in the Linux kernel's KVM lets an attacker with root inside a guest VM read and write the host's memory — but only on AMD hosts with nested virtualization enabled. CVE-2026-99312 was reported through DigitalOcean's bug-bounty program by Oren Yomtov of Terra Security. The upstream fix landed November 2; distribution packages are still catching up, so check your own kernel before you assume you're covered.
What the bug does
When an INIT reset tears down a nested (L2) VM, KVM fails to reset the nested page-table pointer, leaving it aimed at freed memory that can be reallocated. An attacker in the guest can then forge page-table entries and gain read/write access to host RAM. Red Hat's write-up notes the root cause is a page-table pointer missed by the 2022 fix for CVE-2022-3344. Per Red Hat's advisory, the issue affects only AMD hosts running nested KVM; Intel-based KVM and non-nested AMD setups are not affected.
Severity ratings diverge. The Linux kernel CVE team scores it 8.8; Red Hat's preliminary assessment rates it Moderate (5.6) and treats the impact as denial-of-service. In lab testing the researchers read about 127 GB of host memory on a DigitalOcean Droplet (read-only) and wrote 32 bytes into a live host kernel page on their own bare-metal AMD server — so "DoS only" is a floor, not a ceiling.
What's affected
The vulnerable code entered the 6.1 kernel in 2022 and was carried into 5.15.81 and 6.0.11. VirtFusion notes nested virtualization is on by default in the kvm_amd module, so unpatched AMD hosts are exposed unless someone changed the setting.
- Upstream fixes: stable 7.3.5, longterm 6.18.64 and 6.12.118 (commit
c7c288d178cb, "KVM: x86: Unconditionally reset MMU context when emulating INIT", by Paolo Bonzini). - Debian: all releases listed vulnerable; fix in unstable.
- Ubuntu: rated High, no fixed package released at time of writing.
- Red Hat: RHEL 9 and 10 affected, fix deferred; RHEL 8 not affected.
- DigitalOcean live-patched its own fleet and told Droplet customers no action is needed.
Exploitation status
VirtFusion reports no use outside research. No exploit code has been published, and there is no detection guidance for a past attempt. The attacker also needs root inside a guest — network access alone is not enough — which narrows the realistic threat to multi-tenant hosts where an untrusted party can run a guest.
Action checklist
- Check your running kernel against the fixed builds above and patch AMD hosts that run untrusted nested guests first.
- If you can't patch immediately and don't need nesting, disable it and reboot:
# /etc/modprobe.d/kvm.conf
options kvm_amd nested=0
This breaks any VM that itself runs nested VMs. 3. On managed clouds, confirm your provider's patch/live-patch status rather than assuming — Red Hat, Debian and Ubuntu packages were lagging upstream. 4. Treat Intel-only KVM fleets as out of scope for this CVE and prioritize accordingly.
Context
This is the second KVM guest-to-host escape we've covered in recent weeks, after CVE-2026-53359 in the shadow MMU. Nested virtualization keeps producing these: it is complex, AMD-specific code paths get less scrutiny than the common case, and a single stale pointer turns a tenant into a hypervisor. If you don't run nested guests, the cheapest hardening is to turn the feature off.