Anthropic AI Model Finds Linux Kernel Bug That Leads to Remote Root Access
CSO Online, Friday, October 30th, 2026
Anthropic disclosed a zero-click Linux kernel MPTCP exploit, CVE-2026-98432, built with an unreleased Claude model, giving remote root.
Anthropic disclosed a race-condition use-after-free in the Linux kernel's handling of IP options during IPv4 TCP connection establishment, tracked as CVE-2026-98432, which an unauthenticated remote attacker could exploit without user interaction to run commands as root on servers using Multipath TCP.
The exploit was developed with an unreleased Claude research model and leaks kernel memory, defeats KASLR and hijacks control flow.
Anthropic's proof of concept took about 53 minutes from the initial leak to root command execution.
It is part of work evaluating AI vulnerability discovery, and Anthropic disclosed five exploit chains across Chrome, Safari, Android, Windows and Linux.