An unreleased internal model developed by Anthropic has autonomously created an exploit that grants remote root command execution on Linux servers using multipath TCP (MPTCP).
The model was given an initial task and produced working code that leverages CVE-2026-98432, a use-after-free flaw in the Linux kernel network stack. During concurrent TCP processing, one code path can free memory while another continues to reference it, enabling memory corruption.
The attack proceeds in two stages. First the exploit bypasses KASLR kernel address randomization, requiring 48 minutes to locate the necessary memory address. Once the address is obtained, an additional five minutes are needed to hijack control flow and spawn a root shell.
Scoring differences highlight varying risk assessments: the Linux kernel team assigned the vulnerability a CVSS score of 9.8, while Red Hat rated it 5.9, noting the practical difficulties of exploitation in default configurations.
A fix has already been committed to the mainline kernel and all supported stable branches. Administrators are advised to check with their specific distribution vendor for the availability of updated packages.

