Critical Linux kernel CVE-2026-98432 enables remote root via MPTCP; AI bug-hunting floods kernel with reports

CVE-2026-98432 is a use-after-free vulnerability in the Linux kernel's IPv4 TCP handshake code that exists from Linux 4.4 onward and carries a CVSS score of 9.8. The flaw occurs when a server accepts a connection with IPv4 header options in the SYN packet: the kernel transfers those options from the half-open to the full connection without copying, then frees the memory while another CPU core may still access it. Anthropic's Claude model produced a working proof-of-concept that achieves root on servers with Multipath TCP listeners, and scans indicate approximately 2.1 million servers expose MPTCP on port 443 due to Debian's default enablement and Go 1.24's automatic MPTCP socket behavior.

Separately, OpenAI disclosed results from an automated kernel bug-hunting pipeline using an unreleased cyber-focused model. In mid-October, the system identified 2,300 candidate bugs, reproduced 1,700, and generated 948 working local privilege escalation exploits against stock desktop Linux at roughly $1,279 per exploit in API credits. OpenAI states it is coordinating with the kernel community to report findings and submit patches, starting with the most critical issues.

Remote exploitation path and AI-driven vulnerability discovery

• TCP handshake use-after-free: The kernel passes IPv4 header options from the SYN-RECV to ESTABLISHED connection state without copying, then calls kfree() while another core may still read the memory; the four-line fix replaces this with kfree_rcu() to defer freeing until all readers complete.
• MPTCP as attack vector: Multipath TCP listeners are required for the known proof-of-concept; Debian enables MPTCP by default in its kernels, and Go 1.24+ automatically opens MPTCP listening sockets when the kernel supports it, exposing applications like the Caddy web server.
• Exposure metrics: Security researcher Marcus Hutchins found approximately 2.1 million servers reachable via MPTCP on port 443, with active scanning already observed from residential proxies and Russian VPN infrastructure targeting this vulnerability.
• Mitigation without patching: Administrators can detect MPTCP listeners with ss -Mlnp, disable MPTCP via sysctl net.mptcp.enabled=0 (requiring service restart for existing sockets), or apply firewall rules dropping TCP packets with IPv4 options as published by Red Hat and AlmaLinux.
• AI pipeline scale and cost: OpenAI's automated system produced 948 working local privilege escalation exploits in ten days for approximately $1.2 million in API credits; during the same period, over 100 other kernel privilege escalation bugs were patched by human contributors with approximately 300 more under public discussion.

Why it matters

The combination of a remotely exploitable pre-authentication kernel bug with broad MPTCP exposure creates immediate risk for Debian-based servers and Go applications, requiring urgent patching or mitigation across cloud and hosting infrastructure. The demonstrated scale of AI-assisted vulnerability discovery—nearly one thousand working exploits in a single run—signals a fundamental shift in offensive security economics that will strain kernel maintainer capacity and accelerate the arms race between automated attack generation and defensive patching.