CVE-2026-98432: The 9.8 Linux Kernel Flaw Ubuntu and Red Hat Haven’t Patched Yet

CVE-2026-98432 is the kind of vulnerability that makes me stop scrolling. It is a flaw in the Linux kernel itself, the software underneath a huge share of the internet’s servers, it carries a 9.8 out of 10 severity score from the kernel’s own CVE team, and an AI model has already written a working exploit for it. Two of the biggest Linux distributions, Ubuntu and Red Hat Enterprise Linux, had still not shipped a fixed kernel as of October 29.

What CVE-2026-98432 actually is

CVE-2026-98432 is a use-after-free bug in the Linux kernel‘s IPv4 TCP code, and I think the plain-English version matters more than the jargon. When your server accepts a TCP connection, the kernel has to remember a few details about that connection while the handshake finishes. In this case, those details are the IPv4 options attached to the incoming connection, things like Record Route or Timestamp, stored in a field the kernel calls ireq_opt.

The bug is a race condition. One part of the kernel hands those options over to the newly created connection without copying them, while another part of the kernel can still be reading them to build a reply to a retransmitted connection request. If the new connection is freed first, that second part ends up reading memory that has already been released. That is the use-after-free, and in kernel code it is the kind of mistake that can crash a machine or, in the worst case, let an attacker run code with kernel privileges.

A few facts keep this in perspective. The flaw has existed since Linux 4.4, which means it has been sitting in the kernel for a very long time without anyone noticing. IPv6 connections are not affected by this same issue, because the kernel copies IPv6 options instead of sharing them. The upstream fix, by kernel developer Eric Dumazet, is small in concept: free that memory only after a grace period, using kfree_rcu() instead of a plain kfree(). Finding the race was the hard part.

Why the severity scores disagree, and why both sides have a point

If you have been following this story, you have probably seen two very different numbers. The Linux kernel’s own CVE team rates CVE-2026-98432 at 9.8 out of 10, which is about as critical as scores get. Red Hat rates it 5.9 on its CVE page and 8.1 in its advisory file. Amazon Linux scores it 7.0. Ubuntu calls it High priority. So who is right?

Honestly, both sides are describing the same bug from different angles. The 9.8 score is based on a specific attack scenario: an attacker who can reach a server, send a crafted connection request carrying IPv4 options, and win a timing race against the kernel. The lower scores reflect how hard that race is to win in practice. Red Hat’s reasoning, in short, is that the race is difficult to trigger reliably, so the real-world risk is lower than the worst-case score suggests.

That disagreement is normal, but it leaves server owners with a real question: drop-everything emergency, or next-window patch? My read: urgent for internet-facing servers, no panic for a laptop that never listens for inbound connections.

The Anthropic angle: an AI wrote the exploit

The reason this CVE blew up beyond the usual kernel-security circles is Anthropic. On October 27, the company said its AI model had written an exploit for this flaw that achieved remote root command execution against Linux servers running a multipath TCP (MPTCP) listener. The exploit reportedly worked in two stages: leaking a kernel address to defeat address randomization (about 48 minutes of repeatedly triggering the race), then overwriting a callback pointer to run a command as root (about five more minutes).

That is a striking headline, and I want to be careful about what it does and does not prove. It does not mean an AI can now hack any Linux server on its own. The exploit targeted a specific configuration, a server with an MPTCP listener, and Anthropic has not published the exploit code. What it does prove is narrower but still important: AI models are now good enough to turn a subtle race condition into a working proof of concept, given a skilled researcher steering the work. The barrier to weaponizing this class of bug just got lower.

There is also what we do not know, and it matters. As of October 29, there were no reports of this flaw being exploited in the wild, and it was not listed in CISA‘s Known Exploited Vulnerabilities catalog. That could change, which is exactly why the patch status below matters more than the headlines.

Who has patched it, and who has not

This is the part that matters most in practice: waiting for a patch that has not arrived, or assuming you are patched when you are not. Status as of October 29, 2026:

Distribution Status Notes
Debian Fixed Debian 12 (bookworm) ships 6.1.190-1; trixie ships 6.12.113-1
Amazon Linux 2023 Partly fixed kernel6.12 and kernel6.18 packages fixed; the main kernel package was still pending
Ubuntu 22.04, 24.04, 26.04, 26.10 Not fixed yet Priority High, fix in progress, no fixed kernel released
Red Hat Enterprise Linux 8, 9, 10 and OpenShift 4 Not fixed yet Affected; fix deferred at last check
SUSE and Oracle Linux Fixes shipped for some releases Check your vendor advisory for your exact release

Upstream, the fix is in the stable kernel branches, including 6.1.190 and 6.12.113, so if you build your own kernel you can get it now. If you run a vendor kernel, which is what most people do, you are waiting on your distribution, and that wait is the whole story for Ubuntu and Red Hat users right now.

One more detail: Go 1.24 enables MPTCP by default for listening sockets, per the Go security team’s October 29 statement. That does not make every Go program vulnerable, but it widens the set of servers that could run an MPTCP listener, the configuration Anthropic’s exploit targeted.

What to do now: a practical checklist

First, figure out whether this applies to you. If you run a Linux server that accepts connections from the internet, it does. On a desktop or laptop that never listens for inbound connections, this is a patch-when-available situation, not an emergency.

Second, check your kernel and your distribution. Run uname -r to see your kernel version, then check your distribution’s tracker: Ubuntu’s CVE page for CVE-2026-98432, Red Hat’s advisory, or Debian‘s security tracker. Do not assume an apt upgrade fixed it. On Ubuntu, as of October 29, running apt upgrade would not have changed anything, because the fixed packages simply had not been published yet.

Third, prioritize your internet-facing servers, especially any running an MPTCP listener. Those are the systems where the Anthropic exploit scenario is most plausible. If you can, confirm whether MPTCP is actually enabled on your listeners before you spend a weekend rebuilding kernels you do not need.

Fourth, when your distribution ships the fix, install it promptly and reboot into the new kernel. A kernel patch that sits downloaded but unbooted protects nothing, and this is the step people most often forget. Schedule the reboot now, while you are thinking about it.

Fifth, if you run Debian or a fixed Amazon Linux kernel, patch now rather than waiting for a maintenance window to feel convenient. The fix exists for you, and every day you delay is a day the public exploit conversation gets further ahead of your defenses.

Finally, keep an eye on the status pages rather than the headlines. The signal that matters is when Ubuntu’s tracker flips from work in progress to released, and when Red Hat publishes its fixed kernel.

How this differs from the PHP flaw earlier this month

Regular readers may be thinking of the PHP vulnerability covered here recently, and it is worth one sentence of separation: that was a flaw in PHP itself, while this is a flaw in the Linux kernel’s TCP stack. Different layer, different fix, different patch timeline. If you patched PHP, that does nothing for this, and vice versa. Both stories happening in the same month is a good reminder that patching is not a one-time event.

On the Windows side, Microsoft’s November Patch Tuesday fixed 1,094 flaws, including a zero-click Windows Search vulnerability and an Outlook bug already being exploited.

What I am watching next

The next meaningful update will not be another hot take. It will be Ubuntu and Red Hat publishing fixed kernels, and then the slower question of how many servers actually reboot into them. I am also watching whether CISA adds this to its exploited-vulnerabilities catalog, which would be the first hard signal that someone is using it outside a lab. Until then, the honest position is this: a serious, real, 9.8-scored kernel bug with a demonstrated AI-built exploit, no known in-the-wild exploitation yet, and two of the biggest enterprise distributions still working on the fix. That is enough to act on, and not enough to panic about. Patch when your distribution ships it, reboot, and move on.

3 Comments

Leave a Reply

Your email address will not be published. Required fields are marked *