Claude model wrote a no-login Linux root exploit

An unreleased Claude research model wrote a working exploit for CVE-2026-98432, a Linux kernel flaw rated CVSS 9.8 that lets a remote attacker run commands as root with no login and no user action, Anthropic’s Frontier Red Team said in an October 27 post. If your servers run a multipath TCP listener, what you can install depends on the distribution: Debian shipped its fix on October 16, while Ubuntu and Red Hat still had no patch, Golem reported. That may have changed since.

The bug is a race condition in how the kernel handles IPv4 TCP connection handshakes. Red Hat gives the detail. When the kernel processes retransmitted connection requests or socket migrations at the same time, socket options can be freed while they are still in use, a use-after-free. Multipath TCP (MPTCP) is the part that decides whether a server is exposed. It lets a device use several network interfaces at once over a single connection, and each subflow underneath is a regular TCP connection, per the kernel documentation.

Red Hat rates the flaw lower. Its CVSS v3 base score is 8.1 with Attack Complexity High, while cve.org lists 9.8 with Attack Complexity Low. Its advisory says the issue “primarily results in a system crash leading to a Denial of Service (DoS) or potential memory corruption”, which is well short of the root shell Anthropic says its model reached. The same page says mitigation is either not available or doesn’t meet the company’s criteria. Add the missing patch Golem reported, and RHEL users had neither a vendor fix nor a sanctioned workaround at that point.

Tuesday’s post presented the Linux exploit as one of five chains built by the same internal model. Fixes have been slow to follow Anthropic’s reports before. As of October 2, its disclosure dashboard listed 6,157 vulnerabilities reported across 591 open-source projects, adding that “to our knowledge, 516 of these have been patched.”