TL;DR
- what: CVE-2026-99312 is a use-after-free in Linux KVM's AMD nested virtualization code that lets a guest with root access steer the processor to a forged nested page table after an INIT reset.
- impact: Researchers read about 127 GB of a production DigitalOcean host's RAM, which on shared servers can include other tenants' VMs, and wrote 32 chosen bytes into host kernel memory on their own bare-metal server.
- fix: Upstream fixes are in kernels 7.3.5, 6.18.64 and 6.12.118, take effect only after a host reboot, and disabling nested virtualization on kvm_amd is the only mitigation until distribution kernels ship.
- who: Operators of KVM hosts on AMD processors with nested virtualization enabled, which is the kvm_amd default, including hosting providers, private clouds and VPS platforms; Intel hosts are not affected.
A flaw in KVM, the hypervisor built into the Linux kernel, let a researcher read about 127 GB of a cloud server's physical memory from inside a virtual machine they rented on DigitalOcean. On a multi-tenant host, that memory can include other customers' virtual machines. Tracked as CVE-2026-99312, the bug affects only AMD hosts with nested virtualization enabled, and nested virtualization is on by default in the kvm_amd module. Upstream kernel fixes shipped the day before public disclosure. Most distribution kernels are still unpatched.
Oren Yomtov of Terra Security found the flaw and reported it through DigitalOcean's bug bounty program. The provider confirmed the exploit on its own hosts, rated it critical, and live patched its fleet. It told Droplet customers no action was needed on their part. Linux KVM maintainers were notified on October 21 and merged a fix on November 2.
What the Researchers Demonstrated
Two tests show the reach of the bug. In the read test on a DigitalOcean Droplet, the first data returned was the host's kernel version banner, followed by about 127 GB of host RAM mapped read-only into the guest. In the write test, run on Terra's own bare-metal AMD server with no other tenants, the guest wrote 32 chosen bytes into a live page of host kernel memory. A witness module loaded into the host kernel confirmed the change.
Both tests started from root inside a VM the researchers controlled. That is the attacker position required: control of a guest on a vulnerable host. Network access to the host alone is not enough. In a public cloud or VPS environment, though, getting a root guest takes nothing more than a credit card.
⚠️ Cross-tenant exposure — A read primitive covering 127 GB of host memory on a shared server can reach other tenants' guest memory, including keys, credentials and data in use. The write primitive shown on bare metal moves the impact from disclosure to possible host kernel compromise.
Root Cause: A Stale Nested Page Table Pointer
KVM isolates each VM with a nested page table that translates guest memory addresses into real host physical memory. The processor finds that table through a pointer stored for each virtual CPU. When a host allows nested virtualization, a guest can run a VM of its own.
When an INIT reset signal stops the nested VM, KVM fails to restore the outer VM's page table pointer. The pointer is left referencing a page the host kernel has already freed. Once the kernel reuses that page, the guest can fill it with page table entries it chooses, and the processor walks them as if they were the real table. Because forged entries can point at any host physical page, the guest gains a window into host memory it was never assigned. The fix makes KVM reset that memory-management state on every INIT.
The bug was introduced by a security fix. A 2022 patch for CVE-2022-3344, a different nested virtualization flaw, switched a reset path back to the outer VM's control block but missed the page table pointer. The error stayed in the kernel for just under four years. Terra's earlier PatchSlop research found that fixes for other vulnerabilities introduced 3.2% of kernel vulnerabilities.
Scope: Which Hosts Are Exposed
- Affected: KVM hosts on AMD processors with nested virtualization enabled in kvm_amd, which is the default unless an operator has changed it.
- Not affected: Intel KVM hosts, since the vulnerable code exists only in KVM's AMD nested virtualization support.
- Kernel range: the bug entered the 6.1 kernel in 2022 and was backported into 5.15.81 and 6.0.11, so older long-term lines and distribution kernels built on them are in range.
- Several cloud providers Terra surveyed were not vulnerable because they had not enabled nested virtualization on AMD.
The fix lives in the host kernel. Patching guest operating systems changes nothing, and a host keeps running its old kernel until it reboots. Live patching, as DigitalOcean used, is the alternative for fleets that cannot take reboot windows.
Patch Status Across Upstream and Distributions
Upstream, the fix is in stable kernel 7.3.5 and long-term kernels 6.18.64 and 6.12.118, according to the kernel CVE team's announcement. Distribution coverage is far behind:
- Debian: the security tracker lists kernels in all releases as vulnerable, with a fix only in unstable.
- Ubuntu: rated High, kernels listed as vulnerable, no fixed package released yet.
- Red Hat: Enterprise Linux 9 and 10 kernels affected with the fix deferred; Enterprise Linux 8 not affected.
Until a fixed host kernel is installed and running, the only available mitigation is disabling nested virtualization in kvm_amd. That closes this attack path completely, but guests can no longer run VMs of their own, which breaks tenants running Docker Desktop style VMs, Android emulators, CI sandboxes or their own hypervisors. Operators should warn tenants who depend on nested virtualization before changing it. The setting also cannot change while VMs are running, so either route requires a maintenance window.
Severity Ratings Disagree
The Linux kernel CVE team scores CVE-2026-99312 at 8.8, with high impact on confidentiality, integrity and availability. Red Hat's preliminary assessment rates it Moderate at 5.6 and treats the impact as denial of service only. DigitalOcean treated it as critical for its shared cloud. Given the demonstrated 127 GB read and the confirmed write into host kernel memory, the Red Hat rating understates the risk for any multi-tenant AMD host. Teams that triage strictly by vendor score may deprioritize this incorrectly.
Exploitation status — VirtFusion, which is tracking the issue for its operators, reports no use of the flaw outside research. No exploit code has been published, and no guidance exists for detecting a past exploitation attempt. Absence of evidence here reflects missing telemetry, not confirmed safety.
RedEye Assessment
CVE-2026-99312 is the third KVM isolation failure disclosed this year. It follows a 16-year-old flaw in KVM's shadow memory management affecting both Intel and AMD, and the PwnEscape flaw in Linux networking code, which let a guest read host memory in Vercel's live sandboxes in October. The pattern matters for anyone whose security model assumes the hypervisor boundary holds. Any organization that hosts untrusted or semi-trusted workloads on AMD KVM, including internal multi-tenant platforms, should treat this as a priority host patch. Confirm your provider's status if you run on third-party AMD infrastructure, and rate exposure on the demonstrated impact rather than the lowest vendor score.
Originally published on RedEye Threat Intelligence.
Top comments (0)