A virtual machine is supposed to contain an untrusted workload. KVM's boundary is the promise that the guest cannot reach the host kernel, another tenant's memory, or the hardware state that ties them together. CVE-2026-99312 breaks that promise on a specific AMD configuration: a Linux KVM host with nested virtualization enabled, running an affected kernel, and hosting a guest whose owner has root inside the VM.
The flaw is a use-after-free in KVM's nested SVM page-table handling. When a virtual CPU receives an INIT while an L2 guest is running with paging disabled, KVM can return to the outer VM without resetting memory-management state. The outer VM retains a pointer to a nested page table that has already been freed. If the host later reuses that page, the guest can populate it with entries that cause 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. Reading host memory can expose other guests. Writing it weakens the isolation guarantee far beyond a denial-of-service issue. The Linux kernel CNA assigns CVSS 3.1 a score of 8.8. Red Hat later raised its assessment to Important with a 7.8 score after the full impact became public.
What makes the exposure unusual
Nested virtualization is enabled by default in KVM, so a host does not need an explicit nested-workload feature to match the vulnerable condition. AMD hosts are affected; Intel hosts are not. The guest needs root inside its own VM, which is common in infrastructure and research tenants. The host kernel remains the component to patch. Updating a guest's kernel does nothing for this bug.
The bug also shows why CPU-model changes are not a substitute for configuration review. Host passthrough and host-model can give a guest the instructions needed to reach the path. The practical control is to disable nested virtualization on AMD hypervisors or run a fixed host kernel.
Triage the hypervisor, not the application
Start with the CPU vendor and the nested setting:
grep -m1 vendor_id /proc/cpuinfo
cat /sys/module/kvm_amd/parameters/nested
uname -r
An AMD host with 1 or Y in the nested parameter needs a maintenance decision. If the kernel is not fixed, disabling nested virtualization closes the path until a reboot completes. The setting cannot be changed safely while VMs are running. Check kernel command lines and module configuration files as well; a stale kvm_amd.nested=1 can override the intended setting.
The fixed upstream releases listed by the vendor analysis are Linux 7.4-rc2 and stable 7.3.5, 6.18.64, and 6.12.118. Distribution backports differ. A host is not protected until it has rebooted into a fixed kernel and the running version has been verified.
Response priorities
- 1.
Inventory AMD hypervisors with nested virtualization enabled.
- 2.
Disable nested virtualization where tenant workloads allow it, then reboot and verify the runtime parameter.
- 3.
Patch the host kernel to a distribution build that contains the fix.
- 4.
Review high-value and multi-tenant hosts first because the demonstrated impact crosses the guest boundary.
- 5.
Preserve guest and host logs before rebooting if compromise is suspected, while recognizing that no public detection recipe accompanied the disclosure.
No public proof-of-concept or in-the-wild exploitation has been reported in the sources reviewed here. That is not a reason to wait. A configuration that permits cross-tenant memory access is already outside the isolation model most cloud and virtualization teams intend to operate.