In 30 seconds

Security researcher Oren Yomtov says his team found a guest-to-host escape in KVM, the virtual machine manager built into Linux, now tracked as CVE-2026-99312. In plain terms, a rented VM that is supposed to stay locked inside its sandbox found a door into the memory of the physical server underneath it. On DigitalOcean, the team says, an ordinary droplet, which is just DigitalOcean's word for a cloud server, read the hypervisor's kernel banner and then about 127 GB of the host's RAM, the same memory that holds other customers' virtual machines.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
The bug, as described, affects KVM on AMD processors with nested virtualization enabled, and fixes are already listed in mainline Linux and stable releases 7.3.5, 6.18.64 and 6.12.118. The researchers say the attack can also write to host memory, but they demonstrated that part only on their own server, deliberately, to avoid disturbing anyone else. There is no claim so far that anyone exploited this in the wild.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
What happened
The story arrives as a single announcement from Yomtov on X. His team says it found a flaw in KVM, the software that lets one physical machine run many isolated virtual machines, that allows a guest VM to break out of its isolation and access the host's physical memory directly, with both read and write capability.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
To prove the read side, the researchers say they rented a normal droplet on DigitalOcean and used the bug to pull up the hypervisor's kernel banner first, essentially the host's core software introducing itself, and then read roughly 127 GB of the server's physical memory. That memory, they stress, is where every other customer's VM on the machine actually lives.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
For the write side, they say they switched to their own hardware, showing they could modify host memory without risking other people's workloads. According to the same post, the bug requires AMD processors and nested virtualization, meaning VMs running inside other VMs, to be enabled, and patches have already landed in mainline Linux plus the stable 7.3.5, 6.18.64 and 6.12.118 releases.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
Why it matters
The core promise of a public cloud is isolation: your VM cannot see your neighbor's VM. The researchers' claim breaks that promise at the lowest level, because reading a host's physical memory means reading the RAM where other customers' virtual machines on that box actually run. If the claim holds, the cheapest rented machine on the platform could have served as a window into other people's data.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
placeholderorenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
The claimed write capability is arguably the more dangerous half. Reading host memory could expose secrets, but writing to it could, in principle, let an attacker tamper with other tenants' VMs or seize control of the host itself. That half was only demonstrated on the researchers' own server, so real world impact on a provider like DigitalOcean remains untested.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
What to do
If you run KVM yourself on AMD hardware with nested virtualization enabled, the researchers say the fix is already available: update the host kernel to mainline or to stable release 7.3.5, 6.18.64 or 6.12.118. Patch the host, not the guest VMs, because the flaw sits in the hypervisor layer underneath them.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
If you are a cloud customer rather than an operator, you cannot patch this yourself, the hypervisor belongs to the provider, so the practical move is to watch for an advisory or status update from your provider. It is also worth confirming the patched kernel versions with your Linux distribution, since every version number in this story traces to a single announcement.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
Confirmed and still uncertain
What is claimed, all from the researchers' own announcement: the bug exists and is tracked as CVE-2026-99312; it affects KVM on AMD hosts with nested virtualization on; a read of the host kernel banner and about 127 GB of host RAM was demonstrated from an ordinary DigitalOcean droplet; the write was demonstrated on their own server; and fixes have landed in mainline and three stable releases.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
What is not yet established: the available evidence contains no independent confirmation, no vendor advisory and no provider statement, so the CVE details and fix versions rest on one source. We also do not know whether DigitalOcean or other providers have already patched, whether the bug was ever exploited in the wild, when it was introduced, or how many hosts actually run the affected combination of AMD CPUs and nested virtualization.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
What people are saying
The only voice in the current evidence batch is the announcement itself. Yomtov posted: "We found a KVM guest-to-host escape (CVE-2026-99312), exploitable on DigitalOcean", adding that the bug gives a cloud VM read and write access to the host's physical memory and that a full research and exploitation deep dive sits in the thread's comments.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
No reactions, vendor responses or third party commentary appear in this batch, so the wider public conversation around the finding has not been captured here yet.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
Evolution
Unlike security stories that begin with a mysterious incident and fill in later, this one arrived essentially finished. In one post, the researchers say the bug was found, demonstrated on both a live cloud and their own hardware, and already fixed in mainline and stable kernels.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
Patches shipping in multiple stable branches before the public announcement suggests a coordinated, private disclosure to kernel maintainers, with the reveal timed for after the fix. That reading fits the usual pattern, but the announcement itself does not describe how disclosure actually happened.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
Reach
1 independent groups, 1 updates, and 2 engagements.
Technical detail
The bug is tracked as CVE-2026-99312 and is described as a KVM guest-to-host escape. Its preconditions, per the announcement, are specific: the host must run KVM on AMD processors, and nested virtualization, the feature that lets a VM itself host other VMs, must be enabled. The escape is said to grant a guest both read and write access to host physical memory.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
The proof of concept reportedly worked in two stages on DigitalOcean: the guest first read the hypervisor's kernel banner, a sort of boot time identity card for the host's core software, then walked through roughly 127 GB of physical RAM. The researchers point to the comments of their X thread for the full exploitation writeup, which is not part of this evidence batch.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
On the fix side, the announcement lists mainline Linux and stable releases 7.3.5, 6.18.64 and 6.12.118 as patched. Those versions are attributed to the researchers' post rather than to a kernel.org record or vendor advisory in this batch.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
Sources and evidence
This item rests on a single primary source: an X post by researcher Oren Yomtov, published with an attached image, announcing the finding and pointing to a deeper technical writeup in the thread's comments. Every claim in this story traces back to that one post.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
That is a thin base for a story of this size. Nothing in the batch independently verifies the CVE, the fix versions or the DigitalOcean demonstration, and no statement from the provider or kernel maintainers is included. Read it as plausible, detailed first party research that stays unconfirmed until an advisory, a patch record or the full writeup appears.orenyomtovx.com · orenyomtov · xOpen original sourceSee in source list
Sources
- orenyomtovx.com · orenyomtov · x



