· via dev.to (home feed)
KVM AMD nested virtualization flaw CVE-2026-99312 lets guests reach host memory
A use-after-free in KVM's nested AMD page-table handling lets root users inside guest VMs read and write host kernel memory. AMD hosts running with nested virtualization need a patched kernel.

The bug
A use-after-free in KVM's nested AMD page-table code, tracked as CVE-2026-99312, allows a sufficiently privileged guest to break the virtual machine boundary on AMD hosts. According to a technical write-up on dev.to, the flaw affects Linux KVM hosts with nested virtualization enabled. The Linux kernel CNA scored it 8.8 under CVSS 3.1, while Red Hat's assessment, published after the full impact became public, classifies it as Important with a 7.8 score.
How it works
The defect sits in KVM's handling of nested SVM, AMD's nested virtualization extension. When a virtual CPU receives an INIT while an L2 guest is running with paging disabled, KVM returns control to the outer guest without resetting memory-management state. That outer guest then holds a pointer to a nested page table that has already been freed. If the host reuses the page, the guest can fill it with entries that direct the processor to follow mappings into host physical memory.
The published research demonstrated memory reads from inside a cloud VM and writes into host kernel memory on a test machine. Reads can expose data belonging to other tenants; writes go well beyond a denial of service.
Which hosts are exposed
Several conditions define the vulnerable population:
- Only AMD hosts are affected; Intel hosts are not.
- Nested virtualization is enabled by default in KVM, so a host does not need an explicit nested-workload feature to qualify.
- The attacker needs root inside their own VM, which is routine for infrastructure and research tenants.
- Patching the guest kernel has no effect; the host kernel is the component that must be fixed.
The dev.to write-up also notes that CPU-model choices are not a substitute for configuration review. Host passthrough and host-model settings can expose the instructions a guest needs to reach the vulnerable path.
Patching and mitigation
The vendor analysis cited in the write-up lists fixed upstream releases as Linux 7.4-rc2 and the stable 7.3.5, 6.18.64 and 6.12.118 kernels. Distributions backport fixes differently, so a host counts as protected only after it has rebooted into a kernel that carries the fix and the running version has been verified.
Disabling nested virtualization on AMD hypervisors closes the attack path in the meantime. The setting cannot be changed safely while VMs are running, and administrators should also check kernel command lines and module configuration files for a stale kvm_amd.nested=1 value that could override the intended setting.
A practical triage sequence from the write-up:
- List every AMD hypervisor with nested virtualization switched on.
- Disable nested virtualization where tenant workloads allow it, then reboot and confirm the runtime parameter.
- Patch to a distribution kernel build that contains the fix.
- Prioritise high-value and multi-tenant hosts, since the demonstrated impact crosses the guest boundary.
- Preserve host and guest logs before rebooting if compromise is suspected.
Verification starts with three checks: the vendor ID in /proc/cpuinfo, the nested parameter under /sys/module/kvm_amd/parameters, and the running kernel version from uname -r. An AMD host reporting 1 or Y for the nested parameter needs a maintenance decision if its kernel is not yet fixed.
Why it matters
Isolation between tenants is the foundation of shared infrastructure. A guest that can read host memory can potentially see other customers' data, and a guest that can write host kernel memory effectively controls the machine. No public proof of concept or in-the-wild exploitation has been reported in the sources reviewed, but a configuration that permits cross-tenant memory access is not one most operators would deliberately choose. For cloud providers and virtualization teams running AMD hardware with nested virtualization, this is a host-level patch-and-reboot exercise, and the highest priority belongs to the multi-tenant machines where the demonstrated impact matters most.
- #security
- #virtualization
- #kvm
- #linux-kernel
- #amd
- #cloud