Anthropic: Claude Writes Linux Exploit Capable of Remote Command Execution

Anthropic disclosed on November 3 that Claude, following only an initial prompt, wrote a Linux kernel exploit that sends packets to servers with multipath TCP listeners enabled and executes commands with root privileges without credentials or user interaction. The vulnerability, tracked as CVE-2026-98432, has been patched upstream, with fixes backported to all maintained stable kernel branches. Ubuntu's vulnerability tracker shows remediation still in progress for versions 22.04, 24.04, 26.04, and 26.10. The disclosure underscores AI's growing capability to autonomously discover and weaponize kernel-level flaws, raising security implications for enterprises running exposed MPTCP services. Watch for additional vendor advisories and whether the exploit sees in-the-wild use before remaining distributions ship patches.