AI model finds Linux kernel flaw that allows remote root access
An unreleased Anthropic AI model built an attack that can take full control of some Linux servers through a kernel flaw, CSO Online and BleepingComputer report. The attack works on servers running a networking feature called MPTCP. Some Linux versions have fixes, but others are still waiting.
An unreleased Claude model from Anthropic built a working attack on a flaw in the Linux kernel that can let a remote attacker run commands as root, CSO Online and BleepingComputer report. The attack needs no action from the user, CSO Online reports.
The attack works on Linux servers running a listener for Multipath TCP, or MPTCP, the outlets report. MPTCP lets one connection use several network paths at once, BleepingComputer explains. The flaw, tracked as CVE-2026-98432, also affects ordinary TCP connections, BleepingComputer reports. But remote control of a server without an MPTCP listener has not been shown in public, the outlet says.
Anthropic's attack took about 53 minutes to go from the first memory leak to running commands as root, the company said in a blog post, CSO Online reports. The flaw got a severity score of 9.8 out of 10 from the Linux kernel CVE team, BleepingComputer reports. Red Hat rates it 5.9, calling it a possible crash issue that is complex to exploit, the outlet says.
The flaw affects kernels dating back to version 4.4, the outlets report. Fixed kernel releases are out, and Debian, Oracle Linux, SUSE and Amazon Linux have fixes for at least some releases, CSO Online reports. Ubuntu and Red Hat Enterprise Linux had not released patched kernels as of October 30, the outlet says.
MPTCP is not usually on by default, but starting with version 1.24, the Go programming language turns it on for TCP listeners on Linux, BleepingComputer reports. That means some services could expose MPTCP without their administrators knowing, the outlet says. CIQ, a Linux software company, warns that turning off MPTCP alone is not enough, BleepingComputer reports. There are no confirmed reports of the flaw being used in attacks, the outlet says.
What we know: A fix exists in the main Linux kernel, and some distributions have shipped it.
What we don't: When Ubuntu and Red Hat will release patched kernels.
Where outlets agree CSO Online and BleepingComputer report that an unreleased Claude model built a remote root exploit for servers with MPTCP listeners.
Where they differ The cited reports agree on the key facts so far.