
Linux KVM flaw CVE-2026-99312 let AMD-host VMs read 127 GB of host memory
A Linux KVM bug allowed a guest VM on AMD hosts with nested virtualization to read about 127 GB of host memory, with a confirmed write test on bare metal. DigitalOcean patched its fleet and upstream kernels fixed the flaw, but severity assessments differ and detection remains absent. The issue shows how hypervisor-level failures can put multi-tenant isolation at risk even when exploitation requires guest control.
Oren Yomtov of Terra Security reported the KVM issue through DigitalOcean's bug bounty after demonstrating that a guest on an AMD-based host could reach memory outside its own VM. The flaw, CVE-2026-99312, required root inside the guest and affected only AMD hosts with nested virtualization enabled. In one test, the guest read the host kernel banner and about 127 GB of RAM mapped read-only. A separate write test on the researchers' bare-metal AMD server wrote 32 chosen bytes into a live host kernel page.
The evidence trail runs from the bug-bounty report to upstream kernel history. DigitalOcean confirmed the exploit worked on its hosts and rated it critical, while Linux KVM maintainers merged a fix on November 2 after being told on October 21. Terra's analysis says a pointer to the nested page table was left pointing at freed memory after an INIT reset, letting a guest repurpose that page with forged entries that could name any physical host page. The root cause traces to a 2022 patch for CVE-2022-3344.
The disclosure shows a security patch creating a new isolation failure, a pattern seen in hypervisor code where small reset-state omissions become cross-tenant risk. Severity ratings diverge: the Linux kernel CVE team scored the flaw 8.8, citing high confidentiality, integrity and availability impact, while Red Hat's preliminary assessment rated it Moderate at 5.6 and framed it as denial-of-service. That gap matters because cloud operators, not just application defenders, must decide whether a read-only host-memory path is an emergency or a lower-priority kernel update.
The practical exposure is narrow but severe: only AMD KVM hosts with nested virtualization enabled, and attackers must already control a guest. Many surveyed cloud providers were not vulnerable because they had not enabled nested virtualization on AMD. DigitalOcean said Droplet customers need take no action after a live patch, while self-hosted operators must update the host kernel or disable nested virtualization. The missing piece is detection: no public exploit, no confirmed external exploitation and no guidance for finding a past attempt.
Cloud and self-hosted operators running AMD KVM hosts with nested virtualization face a cross-tenant memory exposure if unpatched, while most surveyed providers avoided risk by disabling nested virtualization. The flaw underscores that hypervisor trust boundaries, not just application bugs, can determine tenant data isolation.
Sources (1)
- [1]Original source(https://thehackernews.com/2026/11/new-linux-kvm-flaw-cve-2026-99312-lets.html)