Tech • AI • Robotics • Game

ENFR

Daily Podcast full article

KVM flaw exposes 127GB of memory

A newly disclosed KVM vulnerability, CVE-2026-99312, showed that a guest VM on vulnerable AMD hosts with nested virtualization enabled could cross the line into host memory. Researchers demonstrated roughly 127GB of host RAM exposed on a DigitalOcean Droplet, turning a stale pointer into a sharp reminder that cloud isolation is only as strong as the hypervisor code enforcing it [1].

Generated November 20, 2026 at 6:14 PM≈ 6 min read
AI-generated illustration

The cloud boundary that briefly became porous

The core promise of infrastructure-as-a-service is simple: your virtual machine is yours, your neighbor’s virtual machine is theirs, and the hypervisor keeps those worlds separate. CVE-2026-99312 struck exactly at that promise. Terra Security researcher Oren Yomtov disclosed a guest-to-host KVM escape affecting AMD processors when nested virtualization is enabled, and DigitalOcean later confirmed that the research team demonstrated the read on one of its Droplets .

The most attention-grabbing number is 127GB. According to Terra, the forged mapping on DigitalOcean exposed about 127GB of host RAM in read-only form, enough to search for data that should never be visible from inside a tenant VM . DigitalOcean’s own account says the first proof was the host kernel version banner, followed by about 127GB of host RAM; the write demonstration was performed separately on Terra’s own bare-metal AMD server, where 32 chosen bytes were written into a live host kernel page without putting other customers at risk .

That distinction matters. The cloud-server demonstration proved host-memory read exposure in shared infrastructure; the write primitive was shown in a controlled lab. But for defenders, the operational conclusion is still severe: if a tenant can read host physical memory, data belonging to other tenants can be at risk, and if the bug can be driven toward writable host kernel memory, the path toward deeper compromise becomes much more serious .

What went wrong inside KVM

CVE-2026-99312 sits in KVM’s x86 support for AMD nested SVM, not in ordinary application code and not in the guest operating system. The official vulnerability description says KVM failed to reset the MMU context when emulating an INIT event while a nested guest was running with paging disabled, leaving KVM to resume with stale memory-management state .

In plainer terms, KVM uses page tables to translate what a VM thinks of as physical memory into real host physical memory. On AMD systems, the processor follows a nested page table pointer to decide where guest memory accesses actually land. Terra’s write-up explains that, under the vulnerable sequence, a reset path could leave the VM running with a pointer to a page that KVM had already freed .

Once the kernel reused that freed page, the guest could influence data the processor still treated as a page-table hierarchy. Terra described using guest-controlled entries to point memory accesses at host physical pages, including the host kernel or pages belonging to other VMs . This is why the bug is more alarming than a crash: it crosses the isolation boundary, and it does so from a position many cloud tenants legitimately hold — root inside their own VM .

The bug is limited by several conditions. VirtFusion’s operator guidance says an exposed hypervisor must have an AMD CPU, nested virtualization enabled, a vulnerable host kernel, and an untrusted tenant with root inside a VM . Intel hosts are not affected by this specific path because the vulnerable code path is in KVM’s AMD nested virtualization support; updating a guest VM’s own kernel does not fix the issue because the vulnerable component is the host hypervisor kernel .

Severity: not everyone scored it the same way

The scoring picture is messy because different organizations evaluated the vulnerability from different angles and at different moments in the disclosure cycle. OSV lists CVE-2026-99312 with CVSS 3.1 score 8.8, high severity, with high confidentiality, integrity and availability impact across a changed scope . SUSE’s page also shows the Linux Kernel CNA score as 8.8 while listing SUSE’s own CVSS 3.1 assessment as 7.8, still important and still with high confidentiality, integrity and availability impact .

Ubuntu marked the issue high priority and, as of its 20 November update, listed multiple supported Ubuntu kernel lines as vulnerable, including the main linux package for 26.10, 26.04 LTS, 24.04 LTS and 22.04 LTS . Debian’s tracker showed linux fixed in unstable at version 7.3.5-1 while bookworm, trixie and forky entries remained vulnerable at the time viewed . Oracle Linux assigned the CVE an Important impact rating and listed Oracle Linux 8, 9 and 10 UEK errata dated 19 November 2026 .

These differences do not mean defenders can relax. They reflect packaging states, scoring assumptions and vendor-specific kernel backport practices. The shared message is clearer than the numbers: operators running AMD KVM hypervisors with nested virtualization enabled should assume exposure until they verify the running host kernel contains the fix or nested virtualization has been disabled.

Patch status and mitigations

The upstream fix is direct: reset the MMU context when emulating INIT. OSV lists the CVE as published on 18 November 2026 at 08:42:06Z and identifies fixed Git commits for the stable Linux tree . Terra’s timeline says the fix was merged on 2 November, first tagged in Linux 7.4-rc2 on 9 November, and released on 17 November in stable 7.3.5 and long-term kernels 6.18.64 and 6.12.118 . DigitalOcean says the same fixed versions are the ones operators should target if they run their own AMD KVM hosts .

DigitalOcean said it validated the exploit on its own hosts, prepared a live patch, and rolled the fix across its fleet, telling Droplet customers no action was required on their part . That is reassuring for DigitalOcean tenants, but not a general answer for everyone else. Private-cloud teams, hosting providers, enterprise virtualization clusters and smaller KVM operators have to audit their own hosts.

The immediate mitigation is blunt but effective: disable nested virtualization for kvm_amd. DigitalOcean and Terra both say setting the nested parameter of the kvm_amd module to 0 closes this path when updating is not immediately possible . VirtFusion’s guidance emphasizes that disabling nested virtualization affects workloads that need to run a hypervisor, Android emulator or similar virtualization-dependent software inside a VM, so customers or internal users may need warning before the change .

A subtle operational trap is reboot timing. Installing a fixed kernel package is not the same as running a fixed kernel. VirtFusion notes that a hypervisor remains unprotected until it reboots into the patched kernel and recommends confirming the running version and the nested virtualization setting after maintenance .

Why this one will worry cloud teams

Virtualization bugs are uniquely uncomfortable because they turn normal tenant privileges into cross-tenant questions. Root inside a VM is not supposed to be a privileged position on the host; it is supposed to be an ordinary customer capability. CVE-2026-99312 made that ordinary capability dangerous on a specific class of AMD KVM hosts.

The bug also illustrates a recurring engineering lesson: fixes can introduce future vulnerabilities. Terra and DigitalOcean both note that the flaw traced back to a 2022 fix for an earlier KVM issue, CVE-2022-3344 . The uncomfortable punchline is not that security patches are bad. It is that hypervisor code lives at a level where tiny state-management mistakes can become tenant-isolation failures.

The practical advice is straightforward. Inventory AMD KVM hosts. Check whether nested virtualization is enabled. Confirm the host kernel, not the guest kernel. Apply fixed kernels where available, reboot into them, and keep nested virtualization disabled until the patched kernel is actually running. For providers, also consider whether memory exposure requires customer notification, forensic review or tenant placement analysis.

Apparently, “your memory” briefly had C pointer semantics. The patch restores the intended grammar: a guest pointer must never become a map of the host.

Comments

Be the first to comment.

Sources from the last 72 hours

  1. [1]Virtualization, Responsible Disclosure, and CVE-2026-99312Nov 19, 2026, 1:00 AM
  2. [2]CVE-2026-99312Nov 18, 2026, 9:42 AM
  3. [3]KVM nested SVM (CVE-2026-99312)Nov 19, 2026, 1:00 AM
  4. [4]CVE-2026-99312Nov 19, 2026, 1:00 AM
  5. [5]CVE-2026-99312 Common Vulnerabilities and ExposuresNov 18, 2026, 1:00 AM
  6. [6]CVE-2026-99312Nov 18, 2026, 1:00 AM
  7. [7]CVE-2026-99312: KVM Guest-to-Host EscapeNov 18, 2026, 1:00 AM
  8. [8]CVE-2026-99312Nov 20, 2026, 1:00 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.