Anthropic says its Claude AI model wrote a working cyberattack that lets a stranger take full control of a Linux server over the internet, no password required. The flaw behind it, tracked as CVE-2026-98432, has been present in the Linux kernel since version 4.4 and carries a severity score of 9.8 out of 10.

Response from software vendors has been uneven. Red Hat shipped patches on Sunday, Nov. 1, but Ubuntu—one of the most widely used distributions of the operating system that powers countless servers worldwide—still lists every version from 16.04 to 26.10 as vulnerable in its own security tracker. The Go Security Team says the best fix is to upgrade to a patched kernel, and warns that the bug is not limited to the feature Anthropic's attack used. For anyone running servers, that split means the same 9.8-severity flaw has different fix availability depending on the distribution in use.

How the attack works

Anthropic described the attack on Oct. 27 in a research post covering five targets: Chrome, Safari, Android, Windows, and Linux. Claude produced the Linux exploit after initial prompting, drawing on an unreleased internal model. The attack takes less than an hour to find a hidden spot in the computer's memory, then about five more minutes to run a command with the highest level of access, known as root.

At its core, the bug is a race condition: one part of the system throws away a piece of information while another part is still using it, according to AlmaLinux. A specially crafted message that opens an internet connection can set the condition off, and failed attempts cost the attacker almost nothing. "An enabling property of this race condition is that a lost race only fails the handshake without crashing the whole system," Anthropic wrote, "so a determined attacker can retry indefinitely until they succeed."

Anthropic's version requires a server running a feature called multipath TCP (MPTCP), which allows a single connection to travel over several network routes at once. Programs written in Go have had the feature switched on by default since Go 1.24, according to the Go Security Team. The team cautioned, however, that the exposure is broader than that single feature: "While the demonstrated exploit requires the target system to expose an MPTCP listener, the underlying bug is not specific to MPTCP."

Patching remains incomplete

Red Hat released fixes for RHEL 8, 9 and 10 on Sunday—five days after Anthropic's Oct. 27 research post—including advisory RHSA-2026:93412. Debian, SUSE, Oracle Linux and Amazon Linux have also shipped patches, at least for some versions. For users who cannot update yet, Red Hat published a temporary firewall rule that "reduces exposure to this vulnerability but cannot fully prevent it," the company said in its security bulletin. A complete solution is still under development.

Ubuntu's tracker shows fixes in progress for 22.04, 24.04, 26.04 and 26.10, while versions 20.04, 18.04 and 16.04 are marked only as "Vulnerable," with no release date given for any fix. Until dates appear, the tracker remains the reference for where each Ubuntu release stands.

The episode adds to Anthropic's record in AI-assisted vulnerability research. Its earlier Claude Mythos model helped Mozilla find 271 bugs in Firefox, as Decrypt reported in April. The Linux exploit extends that record from browser code to the kernel itself.