Anthropic: Claude Wrote Linux Exploit Program Allowing Remote Control of Unpatched Servers

Techub News reports that Anthropic announced its AI has written a working attack program that allows attackers to fully control Linux servers over the internet without a password. The vulnerability targeted is CVE-2026-98432, which has existed since Linux version 4.4, with a severity score of 9.8 out of 10. Ubuntu's vulnerability tracker shows all versions from 16.04 to 26.10 are currently affected. Anthropic described this attack in a research article on October 27, covering five targets: Chrome, Safari, Android, Windows, and Linux. Claude, after an initial prompt, used an unreleased internal model to write the Linux exploit. The attack can find the hidden location in computer memory in under an hour, then run commands with the highest privileges (root) in five minutes. Red Hat released patches on Sunday for RHEL 8, 9, and 10. Debian, SUSE, Oracle Linux, and Amazon Linux have also provided fixes for some versions. For users who cannot update yet, Red Hat released a temporary firewall rule but stated it only reduces exposure risk and cannot fully block the attack. Ubuntu's fix is still in progress. (Decrypt)