A twenty-line error entered the Linux kernel on October 2, 2015, passed review by several expert developers and stayed there for eleven years. It was found by a systematic search carried out with artificial intelligence, not a human audit. Eleven days after the discovery, Greg Kroah-Hartman confirmed that the flaw was already fixed in the stable branches 6.12.113, 6.18.56, 7.2.10 and 7.3-rc7The problem is that patching the kernel does not mean your server has it.
The record is CVE-2026-98432, a use-after-free in the IPv4 TCP code that the kernel's own CVE team scored 9.8 out of 10 on October 24. It affects every IPv4 implementation since version 4.4 without exception. IPv6 is out of scope. On October 27, an internal Anthropic research model wrote a working exploit that gives a remote, unauthenticated attacker a root shell, a leap that turns a theoretical entry into a tool available to anyone who knows how to read it.
With that scenario, the figure that matters today is a different one. At the close of this piece, Debian, Oracle Linux, SUSE and Amazon Linux had already published patched kernels for at least some of their versions, and Ubuntu and Red Hat Enterprise Linux had not. An administrator with Ubuntu 26.04 LTS in production faces a window in which the flaw is public, the exploit exists and the official patch has not yet arrived. The reasonable question is not whether to update, but what to do in the meantime with the machines that cannot wait.
Eleven years asleep in the IPv4 code
The error was made in 2015, in version 4.4 of the kernel, and survived eleven years of audits, thousands of reviewers and an industry that prides itself on reviewing the network stack.
The practical consequence of that age is the scope. Any Linux kernel with IPv4 from 4.4 onward carries the problem, and that includes servers, home routers with Linux-based firmware, connected appliances, industrial equipment and most of the container images running in the cloud. The affected versions range from 4.4 to the 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3 branches, according to the advisory published by Hungary's national cybersecurity center, which mirrors the kernel's notice.
Nor does the vector require privileges or user interaction. It is a manipulated TCP connection that triggers the use of already freed memory, and from there to remote code execution is a step that Anthropic's exploit has already taken. Anyone with the machine exposed to the Internet who has not updated has the same problem as an unpatched server had in the weeks of php2shell, a pattern that has repeated itself too often this year.
The real surface is broader than the server label suggests: routers, cameras, industrial equipment and containers share the same network code, and many of them will never see an update again in their useful life.
Who has a patch and who is still waiting
The distribution-by-distribution picture is the most useful part of the advisory and the one no bulletin summarized on Friday. Debian published update DSA-6568-1 with linux 6.12.113-1 for Trixie, and for Bookworm it keeps two paths open, 6.1.190-1 and 6.12.113-1~deb12u1. Oracle Linux, SUSE and Amazon Linux made their move in their support branches. Red Hat keeps the flaw in a state of fix not guaranteed for several versions, a formula that in practice means there is no commitment to a date.
| Distribution |
Status as of October 30 |
Detail |
| Debian |
Fixed |
DSA-6568-1 and two updates for Bookworm |
| SUSE, Oracle, Amazon Linux |
Fixed in supported branches |
Stable kernels released |
| Ubuntu 22.04, 24.04, 26.04, 26.10 |
In preparation |
Kernels planned: 5.15.0-203.213, 6.8.0-149.149, 7.0.0-43.43 and 7.3.0-10.10 |
| Ubuntu 20.04 and earlier |
Vulnerable with no date |
No kernel planned |
| RHEL 8, 9, 10 and OpenShift 4 |
No guaranteed patch |
Remediation status not committed |
The Ubuntu list is the most troubling because it includes the four versions the company considers current. Canonical marks the work as in progress and assigns a kernel number to each branch, which indicates the patch is in the works, but there is no public date. On a 2024 LTS, that leaves thousands of companies facing the decision to wait or move to an alternative kernel, the HWE of a higher version or the distribution's own kernel with partial mitigations.
An Anthropic model wrote the exploit
On October 27, three days after the 9.8, Anthropic reported that one of its research models had produced a working exploit for this flaw, with root shell for an unauthenticated remote attacker. The company runs a program with frontier models used by laboratories and defenders, and this episode fits the line already followed by its most security-focused product. The difference from other announcements is that here the exploit works against kernels that most distributions have not yet replaced.
That is the era shift the advisory hints at. For twenty years, the window between the publication of a CVE and its exploitation was measured in weeks, and time favored whoever patched fast.
With models capable of reading an advisory and writing attack code in hours, that window shrinks to days, and the defensive team now competes against a tool that does not tire. The flaw that lay dormant for eleven years becomes an exploit in three days, and that asymmetry is what security teams will have to accept as normal.
It is worth looking at the other side of the same phenomenon. The AI that found the bug is also the one that makes it possible to audit old code at a scale no human team can reach, and the finding came precisely from a systematic search that reviewed routines nobody had on the radar. The same technology that shortens the reaction margin is the one that uncovered the hole, and no public policy has yet resolved how to share out that double edge. In the ecosystem of agents that analyze code, OpenAI has already had to review its safeguards after similar episodes.
Neither of the two sides solves the immediate problem of the administrator who must decide, today, whether to restart a production service with an exploit already circulating and no official patch in their distribution.
Why is 9.8 not 9.8 for everyone?
The scoring discrepancy is another detail with real consequences. The kernel CVE team rated the flaw at 9.8, while Red Hat, Amazon, Oracle and SUSE place it between 5.9 and 8.1 in their own analyses. The difference is not a calculation error but a difference of judgment about what "remotely exploitable" means when the attacker needs specific network or configuration conditions.
For an administrator, that gap translates into an uncomfortable question. If your distribution's vendor scores the flaw at 6.5, it may not treat it as an emergency and may slot it into the monthly cycle. If the kernel scores it at 9.8, your company's risk committee will want to see immediate action. As long as the two figures coexist, prioritization will depend on which advisory you look at first, and that is exactly what the industry has been trying to solve for years with CVSS 4.0 without fully succeeding.
The technical recommendation is so simple it is almost disappointing. Update the kernel as soon as your distribution publishes it, restart, and do not wait for your vendor's advisory to change color. If you cannot restart, review which services expose IPv4 to untrusted networks and reduce the surface while the patch arrives, because the flaw does not require authentication. And if your fleet is heterogeneous, start with servers running Ubuntu 20.04 or earlier, the only ones that today have neither a date nor a planned kernel.
What remains is the uncomfortable part of any emergency: verifying that the restart actually happened, that the loaded kernel is the new one, and that no automated deployment returned the machine to its previous state the next day.
The era in which AI finds your flaws
The case makes three things clear and none of them is comfortable. The first is that old code nobody audits is still there, at the heart of the network stack, and that tools reviewing thousands of routines per hour are now going to find it. The second is that the time between "there is a flaw" and "there is an exploit" is no longer measured in weeks, and distributions that promise patches in days will fall short at some point. The third is that severity scoring stopped being an objective datum a long time ago.
What comes next is not a new kernel or a global policy. It is a daily race between models that find flaws at scale and teams that update machines at scale, with system administrators in the middle, reading contradictory advisories on a Friday night. The only real advantage the defender retains is the usual one: having an inventory of what is running, knowing what to patch and restarting without asking permission from three committees. If that already works, CVE-2026-98432 is a weekend of work; if not, it is a surprise.