A rented virtual machine can read the memory of the physical machine hosting it, crash the entire machine and, in the lab of the researcher who found the flaw, execute code as root outside its jail. The flaw is called PwnEscape, is registered as CVE-2026-99999 and the Linux kernel CVE team scores it 9.3 out of 10, while Red Hat rates it Important at 8.8.
The path to the entry was short and unusual. Paulos Yibelo, of PWNAI Research, found it in August during the one-million-dollar challenge Vercel launched on its sandboxes, and reported it privately to the kernel security officials on September 4. The accepted patch, signed by Michael S. Tsirkin, of Red Hat, landed in the main branch on October 14 and a day later seven stable kernels were released. The awkward detail is the age of the code: it dates back to 2008, in Linux 2.6.27, so almost any Linux host built since then was born vulnerable.
That is why the urgent fact is not the score, but who has patched. Red Hat, AlmaLinux and Debian published fixed kernels between October 19 and 27, and Ubuntu, with all its supported versions affected, still had no fixed kernel on November 3There is a conceptual trap that explains half the risk: the guest is the system running inside the virtual machine, and the host is the real kernel that runs the hypervisor and to which that guest should only request resources, never read its memory. Updating the guest leaves the host hole where it was.
The checksum that rewrites an already validated header
The flaw is not in the hypervisor, but in how the host kernel processes IPv4 packets arriving from a virtual machine. It is not a KVM bug: it lives in the host's network stack.
Linux allows a packet to travel with the checksum uncalculated, an optimization called checksum offload. The packet carries two hints, where the calculation starts and where the result must be written, and a virtio card allows the guest to pass them to the host. The kernel accepted the position without checking that it fell after the IPv4 header, so the deferred calculation wrote over a header it had already read and taken as valid.
With metadata set incorrectly, the length field went from 5 to 15 in the researcher's example, so the fragmentation code treated a 20-byte header as if it were 60 bytes long and read the next 40 as if they were IP options. That confusion produced three results, described step by step by the researcher and collected in the one-million-dollar challenge:
- Host memory leak, 32 bytes per case and repeatable as long as the guest kept sending packets. Vercel claims it was demonstrated against its production sandboxes.
- Host crash with two packets, which takes down every virtual machine on that machine.
- 64-byte writes in kernel memory chained until a program runs as root on the host, tested only on the researcher's own equipment.
The exploit code is not public and the details of that chain are being held for a second installment, so the practical bar remains high. What is already public is the memory leak and the crash, and neither of them needs an elegant exploit.
What separates the guest from the host and why patching the guest does not help
A provider that updates its virtual machine image believes it has closed the problem and has not touched the door that matters. The vulnerable code runs in the host kernel when processing packets coming from the guest, so the fix has to land there and requires a reboot of the physical machine. Patching the guest protects that guest from the others, never the host that supports them.
Red Hat adds a nuance that broadens the scope. A local attacker can also trigger the flaw, without a virtual machine in between, and the vulnerable code is in all Red Hat kernels, whether or not they run virtual machines. Any affected server goes on the list, although the serious risk is concentrated on those hosting third-party code.
For anyone selling virtual machines, that boundary separates an angry customer from a leak that affects everyone at once. A compromised guest on an unpatched host can read host memory and, through it, that of the other guests on that machine. The system at risk is, in the end, the one running virtual machines belonging to people you do not trust.
The kernel receives more CVEs than it can fix
Who has the kernel patched and who is still waiting
The distribution-by-distribution picture is the most useful part and the one that decides whether a host is exposed this week. The table gathers what each project has published and leaves out products still under investigation, such as NetApp management appliances.
| Distribution |
Status as of November 3 |
Detail |
| Red Hat |
Patched |
RHEL 10 and 9 on October 19 and 20, RHEL 8 on the 24th, RHEL 7 and 6 ELS on the 22nd and 27th; OpenShift 4.12 to 4.22 included |
| AlmaLinux 8, 9 and 10 |
Patched |
Own backport in the production repositories since October 23 |
| Debian 12 and 13 |
Corrected |
Packages 6.1.190-1 and 6.12.113-1 |
| Ubuntu 22.04, 24.04, 26.04 and 26.10 |
Unpatched |
Kernels in preparation; 20.04 and earlier remain vulnerable with no date |
On most distributions the work is already done and what remains is for the host to reboot with the new kernel. Not on Ubuntu, and there a window is opening that is already in its third week. Canonical rates the flaw as Medium priority and assigns a number to each planned kernel, from 5.15.0-203.213 to 7.3.0-10.10, a sign of work in progress with no public date.
Providers that rent out capacity are not in the same boat. Akamai says it has already patched its hypervisors, but its Ubuntu-based nodes were still waiting on Canonical as of October 29. The month is not helping either. CVE-2026-98432, the TCP flaw that Ubuntu and RHEL were still carrying unpatched at the end of October, is a different hole altogether, and the second does not cover the first.
The most common mistake is confusing the installed package with the kernel that is running. `uname -r` answers the second question and that is the only one that counts: a host that updated without rebooting remains exposed, and on a hypervisor that reboot means migrating the virtual machines it hosts first.
MPTCP opens CVE-2026-98432 and Linux remains unpatched on RHEL
The clock that has been running since October 15
The window is not comfortable. Yibelo writes that he expects exploits for weak configurations within weeks and exploits that work with the hypervisor's default configuration in less than a month. Today there is no public exploit, CISA's exploitation assessment in the entry is "none" and the flaw does not appear in the agency's catalog of exploited vulnerabilities.
The risk is concentrated where someone runs third-party code on the same kernel. Cloud and VPS providers, continuous integration runners and platforms that run agents on KVM or Firecracker are the group with the most to lose, because a single malicious tenant can reach the others. Vercel relied on AWS and other partners to coordinate the fix, and almost all the major providers isolate workloads at that layer of the kernel.
The part that cannot be improvised is the inventory. Knowing which hosts run which kernel, which have rebooted and which share a machine with unverified code is what separates a week of work from an incident with a name of its own.
The question worth asking the provider
Anyone renting virtual machines does not need a statement, they need a date. The useful question is whether the kernel of the host supporting your service is patched, on what day that machine was rebooted and whether that happened before October 15. A provider that answers with the guest's kernel has not understood the question, or would rather not answer it.
On the other side, whoever sells capacity has an inventory problem before a technology problem. The window the researcher describes is measured in weeks, and a new audit does not fit in that timeframe; all that fits is having the work done. A flaw born in 2008 that is fixed with a check in six files leaves the feeling that the real problem was not the checksum, but how many hosts had gone eighteen years without anyone looking inside.