Source-reported · Independent verification status is not recorded for this article; claims remain attributed to their sources.
New AMD KVM flaw lets Linux guest read 127 GB of host RAM on nested systems
Researchers have disclosed CVE‑2026‑99312, a KVM vulnerability on AMD hosts with nested virtualization that allowed a Linux virtual machine to access 127 GB of the host’s memory, breaking core assumptions about guest isolation.
A newly disclosed bug in Linux’s KVM virtualization stack has broken one of the basic guarantees of multi-tenant computing by allowing a virtual machine to read large amounts of its host’s memory on certain systems.
The vulnerability, tracked as CVE‑2026‑99312, affects AMD hosts with nested virtualization enabled. Researchers demonstrated that from inside a Linux guest, they could access 127 GB of the host’s RAM, crossing the boundary that should separate a guest from the hypervisor and other workloads.
KVM, the Kernel-based Virtual Machine, underpins many Linux cloud and data-center platforms. It lets operators run many guest operating systems on one physical server under the assumption that each guest is strictly isolated from the host and from other guests. Nested virtualization, where a guest can itself host additional VMs, has become more common in complex development and platform environments.
In the proof-of-concept attack, a malicious or compromised guest with the right exploit could read memory regions that were never intended to be exposed. That opens the way to harvesting encryption keys, authentication tokens, passwords, proprietary code and other sensitive data sitting in the host or in neighboring guests.
Ordinary users won’t see an obvious malfunction. The risk sits with cloud providers, hosting firms and large enterprises that run dense virtualized workloads and rely on software isolation as their main barrier between customers.
For well-resourced attackers, gaining the ability to sit inside a low-privilege VM and quietly inspect host memory is a powerful tool. The scale shown in the research, with 127 GB of host RAM accessed, suggests that significant portions of system state could be exposed before any anomaly is detected.
The flaw also lands at a time when governments and critical industries are pushing more operations into public and hybrid clouds while worrying about technical control. A bug that cuts across hypervisor boundaries reinforces that those boundaries depend on correct code and prompt patching.
In practice, operators will need to identify where they use nested virtualization on AMD hardware, apply kernel and hypervisor updates once available, and consider disabling nested virtualization where patching isn’t immediate. Detecting past exploitation will be difficult because an attack focused on memory reading can leave few traces.
Key indicators to track now are how quickly major Linux distributions and cloud platforms ship fixes for CVE‑2026‑99312, whether follow-on research finds similar weaknesses on other configurations, and whether any evidence emerges that attackers used this path in real-world incidents.