A cloud virtual machine should not be able to inspect the physical server’s memory. A recently patched Linux bug let a security researcher cross that boundary. In a November 19, 2026 technical account, DigitalOcean confirmed that Terra Security researcher Oren Yomtov retrieved about 127 gigabytes of host RAM from a cloud Droplet during authorized testing. The flaw, CVE-2026-99312, affected AMD systems using KVM with nested virtualization enabled. The research demonstration should not be confused with evidence of criminal theft or a widespread customer breach.
Red Hat’s June 2026 official virtualization explainer introduces KVM, QEMU and libvirt. The educational video predates the newly disclosed issue and does not show the exploit. Watch the original Red Hat video.
A guest VM reached memory belonging to its host
Nested virtualization permits one virtual machine to create another virtual machine. Administrators often call the physical host L0, the first guest L1 and an inner guest L2. On affected AMD hosts, a processor reset during nested execution could make KVM retain a reference to a freed nested page-table structure. When the Linux kernel reused the memory, the stale pointer could misdirect virtual memory translation and expose addresses outside the intended guest.
DigitalOcean documented two separate and carefully controlled demonstrations. On a shared-cloud Droplet, the researcher read a host-kernel version banner and then approximately 127 GB of host RAM, read-only. On a separate bare-metal AMD test host with no other customers attached, he demonstrated a 32-byte modification of a live host-kernel memory page. The second experiment showed that writing memory was possible without attempting it on a shared production host.
AMD nested virtualization determines the affected scope
Red Hat’s CVE assessment narrows exposure to AMD-based KVM with nested Secure Virtual Machine (SVM) operation under specific conditions. Intel-based KVM hosts and AMD deployments without nested virtualization are outside the described vulnerability. Updating software inside a guest alone is insufficient; the failure occurs in the host’s Linux hypervisor kernel. Providers and self-hosted administrators therefore have different responsibilities.
IBM Technology’s earlier official hypervisor video explains why physical servers can host many logically isolated operating systems. It provides background on the security boundary, rather than footage of CVE-2026-99312. Watch IBM’s original video.
Kernel patches shipped November 17
According to DigitalOcean, the correction was first tagged in Linux 7.4-rc2 on November 9 and appeared in 7.3.5, 6.18.64 and 6.12.118 on November 17. Enterprise Linux distributors may backport security patches without changing their product’s headline kernel version. Operators should follow their distribution’s package advisories and restart or reboot hosts as required. DigitalOcean says it validated the report, prepared a live kernel patch and addressed the affected pathway across its Droplet infrastructure.
For self-managed AMD KVM servers awaiting a corrected kernel, Red Hat documents turning off nested virtualization with the kvm_amd module’s nested=0 setting. The mitigation can disrupt workloads relying on VMs inside VMs, so changes require inventory and maintenance planning. Administrators can check their host configuration rather than assume every AMD processor is exposed.
Why one exploit matters beyond one provider
A hypervisor is supposed to prevent neighboring tenants from reading each other’s data, even when they share DRAM and CPU cores. The 127 GB read was an observed research result, not an industry-wide estimate of compromised memory. Earlier BitcoinVersus.Tech reporting covers hypervisor architecture, virtual-machine isolation, Talos Linux virtualization research, and a separate KVM escape disclosure. Similar attack classes do not imply identical software defects.
Future follow-ups should verify host-patch completion, nested-virtualization settings and evidence of actual exploitation. A proof of concept demonstrates a boundary failure; it does not, by itself, establish that outside attackers used the vulnerability.
Sources: DigitalOcean’s November 19 technical disclosure; Red Hat security assessment; official CVE record; Ubuntu kernel advisory. Fixed-version availability varies across vendors.
We volunteer daily to ensure the credibility of the information on this platform is Verifiably True. If you would like to support to help further secure the integrity of our research initiative, please scan the QR code below.


Leave a Reply