> status: mostly fine. patch anyway.practical AI security · est. 2026

Calm down. · Builders

Calm down: an AI turned an 11-year-old Linux TCP bug into root, and the fix is still a reboot

By alkizz ·

On Oct 27, Anthropic published a post saying an unreleased internal research model, "after initial prompting," wrote an exploit that lets a remote attacker send packets to Linux servers with a multipath TCP listener and run commands as root, with no credentials and no user interaction. The bug is CVE-2026-98432, a use-after-free in the kernel's IPv4 TCP handshake code, and the kernel's CVE team scores it 9.8, so you can guess how the rest of the week went. It's real, but it's already fixed upstream, nobody has published a working exploit and nobody has reported attacks. Calm down.

In Anthropic's proof of concept it took 48 minutes to leak a kernel pointer and another five to run a command as root, and a lost race just fails the handshake rather than crashing the box. Anthropic hasn't released the code.

Claimed vs. confirmed

What the doom version says What the sources say
An AI can now root any Linux server Root has only been shown in Anthropic's own proof of concept, against servers with a multipath TCP (MPTCP) listener, and nobody has independently reproduced it
It's being exploited right now It's not in CISA's KEV catalog (the Oct 28 release), there's no validated public exploit, and vendors, the kernel CVE team and Anthropic report no attacks
It's a 9.8, so drop everything The 9.8 was added Oct 24, nine days after the CVE went public. Red Hat's CVE page, SUSE and Oracle rate it 8.1 and Amazon 7.0, and Red Hat says its score reflects "the high attack complexity of winning the race"
Just turn off MPTCP MPTCP only shows up in Anthropic's post. The CVE record and the fix describe plain IPv4 TCP, no vendor lists switching MPTCP off as a mitigation, and Red Hat says no available mitigation meets its criteria
It was fixed Oct 5, so you're fine Oct 5 is when it landed in the kernel's networking tree. It reached mainline Oct 9 and shipped in 7.3-rc7 and 7.3, and what matters is your distro's kernel, which on Ubuntu and RHEL doesn't exist yet

The boring lesson

The part worth taking seriously isn't this one bug. The flaw came in with a 2015 commit and first shipped in Linux 4.4, so it sat in the TCP handshake code for about 11 years. Florian Roth made the point well: for most of that time, attackers were about as limited as the people writing the code, and AI is breaking that. A model that can be pointed at a networking stack and come back with a working root exploit is chipping away at that limit, and pretending otherwise is its own kind of denial.

But look at who got there first this time. The fix credits an Anthropic researcher as the reporter, and it was merged into the kernel weeks before the post went up, which is about the best argument you'll find for defenders having the same tools attackers are going to have. The fix itself is as boring as it gets.

What to do:

Sources

← All briefs