The Guest Read the Host. All 127 Gigabytes.

The Guest Read the Host. All 127 Gigabytes.

A virtual machine is supposed to be a box. Whatever runs inside it stays inside it — that is the whole promise of the cloud. This week, researchers showed that promise breaking on AMD hosts: from inside a virtual machine of their own on DigitalOcean, they read about 127 gigabytes of the host server’s memory. On a shared server, that memory can hold other customers’ VMs. The flaw is tracked as CVE-2026-99312, and it affects only hosts with AMD processors and nested virtualization enabled.

The flaw was found by Oren Yomtov, a researcher at the security firm Terra Security, who reported it through the provider’s bug bounty program; the provider confirmed it and treated it as critical. The Linux KVM maintainers were told on October 21 and merged a fix on November 2. Fixed kernels shipped upstream on Tuesday, the day before publication, and the provider live-patched its own fleet, telling Droplet customers: “No action needed on your part.”

How the researchers proved it

On one of its Droplets, the host’s kernel version banner came back first, then about 127 GB of the host’s RAM, mapped read-only. The write test ran on the researchers’ own bare-metal AMD server, with no other customers attached. There, the guest wrote 32 chosen bytes into a live page of the host kernel’s memory, confirmed by a witness module in that kernel.

Both tests started with root access inside a VM controlled by the researchers. An attacker would need the same position: control of a guest on a vulnerable host, not just network access to it. On a public cloud, renting a VM is exactly how you get there. And the window between disclosure and exploitation can be short: the OpenSSH flaw earlier this month saw its first exploitation attempt within nine hours of disclosure.

Diagram: the KVM flaw CVE-2026-99312 and its fix — a freed page-table pointer lets a guest forge memory-table entries and read 127 GB of host memory; the fix is to patch the host kernel to stable 7.3.5 or LTS 6.18.64 / 6.12.118, reboot the host, or turn nested virtualization off with options kvm_amd nested=0.

How it works

KVM confines each VM with a table — the nested page table, which translates VM addresses into the host’s real memory — found through a pointer kept for each virtual CPU. When a reset signal called INIT stops a nested VM, KVM skips restoring the outer VM’s pointer, leaving it pointing to a page of memory the kernel has already freed. Once the kernel reused that freed page, the guest filled it with entries of its own choosing, and the processor followed them as the real table. The forged entries could name any physical page on the host — which is how about 127 GB of the server’s RAM came into reach of a guest that was never given it.

The flaw was created by a security fix. A 2022 patch for another nested-virtualization bug, CVE-2022-3344, missed the page-table pointer, and the error then sat in the kernel for just under four years. In earlier work called PatchSlop, the same team found one fix introduced 3.2% of kernel vulnerabilities.

Who is affected — and how bad is it?

Only AMD hosts with nested virtualization enabled are in range. Intel-based KVM hosts are not affected, because the vulnerable code exists only in KVM’s support for nested virtualization on AMD. Many providers surveyed were not vulnerable — nested virtualization was never enabled. But nested virtualization is on by default in the kernel’s kvm_amd module, so an unpatched host is exposed unless the setting was deliberately changed.

Severity ratings disagree. The Linux kernel’s CVE team scores the flaw 8.8 out of 10, while Red Hat’s preliminary assessment rates it as Moderate at 5.6 and treats the impact as denial-of-service only. The provider treated it as critical for its own shared cloud.

There are no reports of the flaw being used outside research, according to VirtFusion, which is tracking the issue for operators. No exploit code has been published, and no guidance exists for detecting a past attempt. If someone had used it, there is currently no way to tell.

What is still vulnerable

The fix lives in the host kernel: stable 7.3.5 and long-term kernels 6.18.64 and 6.12.118. The code dates to the 6.1 kernel in 2022, carried into the 5.15.81 and 6.0.11 releases. As of publication, Debian lists the kernels in all of its releases as vulnerable, with a fix only in unstable, and Ubuntu rates the flaw High and lists its kernels as vulnerable with no fixed package released. Red Hat lists the Enterprise Linux 9 and 10 kernels as affected with the fix deferred, and the Enterprise Linux 8 kernel as not affected. Distribution kernels mostly do not have the fix yet — the second kernel flaw in a month to hand an attacker the keys, after a 9.8 TCP flaw needing no login.

Diagram: who is exposed to CVE-2026-99312 — an AMD host with nested virtualization on and a guest with root access leads to one question, is the host kernel patched; patched and rebooted means safe, patch pending means turn nested virtualization off, unpatched with nested virtualization on means the host is exposed; the destination is patch and reboot.

What to do now

Update the host kernel — guest OS updates change nothing; a host keeps its old kernel until rebooted. Until a fixed host kernel is installed and the host rebooted, the only available protection is to turn nested virtualization off. That closes this path completely — but warn tenants first, since their VMs can then no longer run VMs of their own. Check the setting with cat /sys/module/kvm_amd/parameters/nested — a value of 1 or Y means nested virtualization is on — and if the host cannot be updated yet, add options kvm_amd nested=0 to the host’s module configuration, then reboot. Confirm with uname -r afterwards.

Patch the host, reboot it, and check that nested setting today.

Source: The Hacker News.

Comments

This site uses Akismet to reduce spam. Learn how your comment data is processed.