Anthropic says Claude found an 11-year-old Linux kernel flaw

Anthropic has published a security research report describing how one of its unreleased internal Claude models discovered a severe vulnerability that had gone unnoticed in the Linux kernel for 11 years. The flaw is tracked as CVE-2026-98432 with a CVSS score of 9.8, in the highest-risk category, and it stems from a use-after-free (UAF) defect in the IPv4 TCP handshake path. The problematic code was submitted in October 2015 and shipped with the Linux 4.4 kernel, where a CPU can read memory that another CPU has already freed while handling TCP connection retransmissions. An attacker needs no password and no user interaction; a network packet is enough to trigger it and obtain remote Root code execution. The kernel CVE team classifies it as a critical issue that is network-reachable, requires no privileges, and requires no interaction.

The research also describes how the Claude model went beyond detection and generated a working two-stage exploit. The first stage spends roughly 48 minutes defeating kernel address randomization, and the second stage takes only 5 minutes to execute arbitrary commands at Root privilege, for a total of about 53 minutes. Importantly, a failed attempt does not crash the target, so an attacker can retry repeatedly. Upstream patching was completed on October 15, with fixes released for 5.10.272, 6.1.190, 6.12.113, 6.18.56, 7.2.10, and 7.3-rc7. Debian, Red Hat, Amazon Linux, and other distributions have since rolled out updates, but Ubuntu 22.04 LTS, 24.04 LTS, 26.04 LTS, and 26.10 remained flagged as affected as of November 1, with fixes still in development.