Critical CVE-2026-99312 Vulnerability Exposes Host Memory via Linux KVM on AMD Systems

A severe vulnerability dubbed CVE-2026-99312 has been discovered in Linux’s KVM hypervisor, allowing guest virtual machines running on AMD-based hosts with nested virtualization enabled to access large swathes of host memory, including that of other tenants. The bug was identified and responsibly reported by researchers at Terra Security after testing on one cloud provider’s infrastructure. Fixes have been released, but many systems remain exposed.

How the Exploit Works

The flaw affects AMD hosts where nested virtualization is turned on. In testing, researchers were able to read about 127 GB of host RAM from inside a guest VM, despite not being granted those permissions—effectively a breach of the isolation KVM is meant to enforce. Further, they demonstrated write access on a bare-metal test machine by altering 32 bytes in live kernel memory using a malicious nested VM, with the change verified by a kernel witness module. Root privileges inside the guest VM are required to exploit this bug; mere network access is insufficient.

Root Cause and Fix

The vulnerability arises from a pointer for nested page tables that isn’t reset during certain INIT reset paths. When a nested VM is reset, the outer VM pointer that refers to the page tables isn’t restored. That pointer ends up referencing freed kernel memory; if that memory gets reused, a malicious guest can write its own page-table entries, potentially mapping arbitrary physical memory pages.

The faulty behavior was introduced in 2022 by a patch intended to correct another nested virtualization issue (CVE-2022-3344). That fix missed this pointer handling path, allowing the bug to linger in the kernel for nearly four years across multiple long-term kernel versions.

Impact, Severity, and Scope

Only AMD KVM hosts with nested virtualization enabled are vulnerable. Intel systems aren’t affected. Some cloud providers are safe if they’ve left nested virtualization off. At least one provider—where the bug was discovered—rolled out live patches to their fleet, indicating no action is required by end users. For operators managing their own AMD-based hosts, patching the host kernel or disabling nested virtualization are the only mitigations.

Severity assessments vary: the Linux kernel team ranked the flaw with a CVSS of 8.8, citing serious risks to confidentiality, integrity, and availability. In contrast, one major vendor’s initial rating was 5.6 and viewed the issue primarily as a denial-of-service threat. To date there’s no evidence the flaw has been used in real-world attacks, nor has published exploit code been made available.

Mitigation and Status of Patches

The vulnerability has been patched upstream. Fixes are included in mainline kernel version 7.3.5 and in long-term maintenance releases 6.18.64 and 6.12.118. Systems running kernels based on older branches may still be vulnerable. For example, Debian’s stable and Ubuntu’s supported releases have not yet distributed patched packages widely, and some enterprise Linux versions are either affected or awaiting fixes.

Until the patch is applied and the host rebooted, the only guaranteed workaround is disabling nested virtualization. Operators are advised to check if nested virtualization is active (via the kernel module parameter “nested”), and if necessary, set it to off in module configuration. Note rebooting is required both after disabling the feature and after installing a fixed kernel for protections to take effect.

The vulnerability is one of several KVM isolation breakdowns discovered this year, following flaws in shadow memory management and sandbox escapes in managed hosting environments.

What this means: CVE-2026-99312 reminds us that seemingly minor omissions in memory‐management logic can lead to mass exposure in cloud and virtualization environments. For cloud tenants, it’s a red flag about trusting isolation guarantees—especially on shared infrastructure. Operators should urgently audit their AMD hosts, identify if nested virtualization is enabled, and apply updates or disable that feature. In future, careful review of reset paths and pointer hygiene in nested virtualization code will be essential to avoid similar catastrophic oversights.