- Terra Security found that a Linux KVM flaw, CVE-2026-99312, let a guest VM read about 127 GB of a DigitalOcean host's memory.
- Only AMD hosts with nested virtualization enabled are affected. Fixed kernels shipped on November 17, and turning nested virtualization off also closes the path.
- Teams running their own KVM hosts should check CPU vendor, the nested setting and kernel version. Cloud customers should ask their provider about patching.
A cloud's promise rests on one pointer
Cloud computing sells a simple promise. Your virtual machine sees only its own memory, even when the physical server also hosts strangers. Terra Security reports a break in that promise inside KVM, the hypervisor that ships within the Linux kernel. The flaw is tracked as CVE-2026-99312.
The lesson here is that a boundary this important can hang on a single field. A patch written for an older bug can loosen it without anyone noticing.
What the researchers showed
Terra's researchers ran the first test on a DigitalOcean droplet. From inside their guest VM, they built a fake memory map that exposed roughly 127 GB of the host's RAM for reading only. Scanning that memory turned up the host's kernel version banner, which Terra took as proof. The Hacker News names the finder as Oren Yomtov of Terra Security.
The write test ran on Terra's own server, so no one else's infrastructure was at risk. There, a guest wrote 32 chosen bytes into a host kernel page. A monitoring module the researchers had loaded confirmed the change, according to The Hacker News.
In each test, the researchers already had root control of the guest VM. An attacker would need the same foothold. Network access alone is not enough.
How the flaw works
A guest VM sees memory that starts at address zero. KVM translates those addresses into real host memory through a table called the nested page table. KVM keeps the table out of the guest's reach. The processor reads it through a pointer stored for each virtual CPU.
Some clouds let a VM run another VM inside it. While the inner VM runs, the processor uses the inner VM's pointer. When the inner VM stops, KVM normally restores the outer VM's pointer.
One way of stopping it, a reset signal called INIT, skips that step. The outer VM resumes with its old pointer, and the page it names has already been freed.
The kernel then reuses that freed page as a bottom-level table filled with entries for the guest's own memory. The processor still treats the page as the top of the map. It reads one of the guest's ordinary pages as the next level down.
The guest controls that page, so it can write a map of its own. Its entries can name any host page, including the host kernel or other customers' VMs. The fix makes KVM reset that state on every INIT, according to The Hacker News.
A security fix created the bug
Terra says an earlier patch caused the flaw. That patch, for CVE-2022-3344, repaired a different nested-virtualization bug. It pointed a reset path back at the outer VM's control block but overlooked the page-table pointer. The gap stayed in the kernel for nearly four years.
Terra also cites its earlier PatchSlop research. In that work it found that 3.2% of kernel CVEs came from the fix for some other CVE. That is Terra's own measure. It does not show a trend across the industry.
Who is exposed, and how serious it is
Terra says the flaw needs AMD processors with nested virtualization enabled. The Hacker News reports that Intel-based KVM hosts are not affected. Terra adds that many of the cloud providers it surveyed escaped because they never turned on nested virtualization for AMD.
DigitalOcean applied a live patch to its own servers. It told droplet customers: "No action needed on your part."
VirtFusion, a virtualization vendor tracking the issue, reports no sign of use outside research. The Hacker News says no exploit code is public. It also says there is no guidance for detecting a past attempt, so teams cannot easily confirm what did not happen.
Severity ratings disagree. The Linux kernel CVE team scores it 8.8 out of 10. Red Hat's early assessment is lower: 5.6, labelled Moderate, on the view that the harm is limited to a service outage. DigitalOcean treated it as critical for its shared cloud. Neither source explains the gap. One plausible reading is that a shared cloud, where guests are strangers, has more to lose from a memory read than a service outage.
What to ask your team
Terra gives this timeline. It reported the flaw to the KVM maintainers on October 21. A fix was merged on November 2. Stable kernel 7.3.5 and long-term kernels 6.18.64 and 6.12.118 shipped on November 17.
Start with four questions. Do we run our own KVM hosts, and are they AMD? Is nested virtualization on, and does anyone need it? Who owns the host kernel update, and by when? If we rent capacity, has our provider confirmed it is patched or not exposed?
For hosts that cannot take the kernel update yet, Terra's advice is to set the kvm_amd module's nested option to 0. In Terra's words, this switches nested virtualization off and "closes this path completely." That suits only hosts where no guest depends on nested virtualization, since turning it off would stop those guests from running their own VMs.
The wider point is for anyone who buys isolation as a service. Ask how a vendor tests its own patches, not only how fast it ships them. A fix is new code, and this one stayed open for four years.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Terra Security.





