A nasty Linux kernel bug lets attackers take over servers, and Anthropic says its AI wrote the exploit

Key points

• Anthropic said on October 27 that its own AI model wrote a working exploit for a Linux kernel bug, CVE-2026-98432, that lets a remote attacker take full control of a server without a password.
• The flaw sits in the kernel's TCP code, the piece of Linux that handles every incoming network connection, and the Linux kernel team scored it 9.8 out of 10 for severity.
• Kernel maintainer Eric Dumazet committed the fix, and it landed in Linus Torvalds' tree on October 5.
• Ubuntu and Red Hat Enterprise Linux ship the affected code, so patches must flow through each vendor before most servers see them.
• IPv4 is the vulnerable path; the bug does not hit IPv6 the same way because that code already copies the relevant data.

A bug deep inside Linux lets an attacker run commands as the top-level "root" user on a vulnerable server, no login required. That's as bad as server bugs get. What makes this one worth reading past the usual patch advisory is that Anthropic says one of its AI models produced a working exploit.

Tracked as CVE-2026-98432 and first reported by The Hacker News, this flaw carries a 9.8 out of 10 score on the industry severity scale from the Linux kernel's own security team.

What is actually broken?

When a computer on the internet knocks on a Linux server's door, the kernel sets up a small record to track the handshake. In some situations two parts of the kernel use that record at the same moment, and one frees the memory while the other is still reading it. Reading freed memory is a classic "use-after-free" bug. On a server, it can be bent into letting an attacker run their own code.

The failure mode is a race between a retransmitted connection attempt and the code building the reply. On a busy server with many cores, that race isn't exotic. It happens under load.

We've covered this class of flaw repeatedly: our 23 September story on Ubuntu shipping a kernel use-after-free found the Unix socket garbage collector still unpatched in affected LTS releases weeks after the upstream fix landed. History rhymes.

Eric Dumazet's fix delays freeing the memory until it's safe. It landed in Torvalds' tree on October 5.

Who is exposed?

Anything running a mainline Linux kernel that speaks IPv4, which covers almost every Linux server on the internet. Ubuntu and Red Hat Enterprise Linux are named, and in practice that pulls in most cloud images on AWS and Azure, plus GCE VMs and the Linux boxes sitting in company data centres.

IPv6-only services aren't hit. That code path copies the relevant data instead of handing off a pointer, so the race doesn't occur. Mixed stacks still need patching because the IPv4 code is still loaded.

Cloud customers don't patch the kernel themselves on managed services. The managed database, the serverless runtime: the provider handles those rollouts. For your own EC2 instances and self-managed VMs, it's on you, through whichever distro's update channel you use.

The AI-wrote-the-exploit part

Anthropic's claim, made on October 27, is that its model produced a remote root exploit for this bug. Vendor PR about AI writing exploits usually deserves a slow blink. The question the post-mortem needs to answer is whether the model built the exploit from scratch or stitched together public kernel exploitation patterns around a fresh bug. Both are notable. They're not the same thing.

The operational point holds regardless. The window between a patch appearing in a public git tree and a working exploit existing is shrinking, and automated tooling is part of why.

Should you worry?

If you run Linux servers, pull the kernel updates your distribution ships and reboot. Running a laptop or home machine on Linux, accept the next kernel update when it arrives. Your organisation runs on a cloud provider's managed service, you don't need to act, but it's fair to ask your provider when their fleet is fully patched.

Update the kernel. Reboot. Move on.