Anthropic has published a zero-click chain that turns a Linux kernel bug into a root shell without any user interaction, and the exploit was written by a research model the company has not released.
The target is CVE-2026-98432, a race-condition-induced use-after-free in the kernel’s networking stack. Servers running Multipath TCP listeners form the exposed population.
The primitive comes from how Linux handles IP options during IPv4 connection setup. Ownership of those options shifts to the freshly created socket from the pending request, and no duplicate is made. Should a second core assemble a SYNACK from that block while the child socket is being destroyed, the kernel reads freed memory.
Anthropic’s proof of concept moved from leaking kernel memory to defeating KASLR, hijacking control flow and finally running an attacker-chosen shell command. The whole run took roughly 53 minutes, the company said in a blog post.
The work belongs to a wider evaluation in which the same internal model produced five exploit chains spanning Chrome, Safari, Android, Windows and Linux, showing how separate weaknesses stack into functional attacks.
The bug is scored critical at 9.8, and eight branches carry it, back to 4.4. Kernel teams published 5.10.272, 5.15.223, 6.6.159, 6.1.190, 6.12.113, 6.18.56, 7.2.10 and 7.3-rc7. A memory-management correction upstream now waits out an RCU grace period before releasing the objects in question, which leaves concurrent readers enough slack to wrap up.
Debian, Oracle Linux, SUSE and Amazon Linux ship updated builds for at least some releases, while Ubuntu and RHEL users still lacked one on October 30.
Administrators should check kernel versions, apply updates and work out whether MPTCP is reachable at all. Hosts that never load that path have nothing to exploit, which narrows exposure while vendor builds arrive.
