An unreleased Claude model built by Anthropic has found a critical flaw in the Linux kernel, the core software that runs most of the internet's servers, and demonstrated how it could be used to seize full control of an affected machine. The bug, tracked as CVE-2026-98432, carries a severity score of 9.8 out of 10, according to How-To Geek, which first reported the finding.
The flaw sits in the kernel's networking code, specifically in how it handles Multipath TCP, or MPTCP, a protocol that lets a device stitch together multiple connections, say Wi-Fi and mobile data, into one session. It's what's known as a use-after-free bug: the kernel hands connection settings to a new socket, then clears that memory too early while another process is still trying to read it. That mismatch is what an attacker can abuse.
Anthropic's proof-of-concept didn't just flag the bug. It built a working chain that leaks kernel memory, defeats a protection called KASLR that normally hides where critical code lives, and runs an attacker's chosen command with root privileges, the highest level of access on a Linux system. No clicks, no password and no file opening required on the target's end. How-To Geek reported that the full chain, from first leak to root access, took the model about 53 minutes to complete.
That's a notable jump in what AI systems are being shown capable of in security research. Finding a kernel-level memory bug is one thing. Turning it into a reliable, working exploit end to end is a different and considerably harder task, one that's traditionally needed a skilled human researcher.
Which systems are exposed
Eight Linux kernel branches are affected, with the bug present as far back as version 4.4, according to How-To Geek's report. Fixes have landed in kernel versions 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3-rc7.
But a kernel fix upstream doesn't mean every server is protected yet. Debian, Oracle Linux, SUSE and Amazon Linux have shipped patched packages for some of their releases, per the report. Ubuntu and Red Hat Enterprise Linux had not shipped fixes at the time the story was published.
That gap matters because most servers don't run a raw kernel release, they run whatever version their distribution packages and ships on its own schedule. Until a distro pushes its patch, systems using MPTCP listeners stay exposed regardless of what the upstream kernel project has already fixed.
Administrators are advised to check which kernel version their servers are running, confirm whether MPTCP listeners are enabled, and apply the patched kernel as soon as their distribution makes one available. Anthropic has not said whether or when it plans to release the model that found the bug, and the company has not detailed what other vulnerabilities, if any, the same research effort has turned up.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.