Security teams are scrambling to patch a critical Linux TCP vulnerability tracked as CVE-2026-98432, which allows remote, unauthenticated attackers to execute commands as root on affected servers. The kernel flaw, disclosed in late October 2026, carries a maximum CVSS score of 9.8 and affects code running since Linux 4.4.
Key points
- The flaw sits in the IPv4 TCP handshake handler and triggers before any user authentication or application logic runs.
- Anthropic confirmed that its Claude model generated a working proof-of-concept exploit granting root access to machines running multipath TCP listeners.
- Upstream stable kernels contain the fix, but major distributions including Ubuntu and Red Hat Enterprise Linux are still preparing updates.
- Administrators can mitigate exposure immediately by filtering packets with IPv4 options or disabling multipath TCP sockets.
How the memory corruption occurs
The issue sits in the core networking stack that processes inbound IPv4 connections. When a remote client opens a connection, the Linux kernel logs the initial handshake state inside a lightweight structure called a request socket. Once the three-way handshake finishes, the kernel creates a standard socket to hand over to user applications.
If the incoming SYN packet includes IPv4 header options, such as Timestamp or Record Route, the kernel saves them to mirror in its replies. The flaw occurred because the kernel transferred references to those saved options from the request socket to the new socket without copying the underlying memory. When the connection closed, the kernel freed that memory immediately, even while another processor core was still reading the options to build a handshake reply packet.
Because the kernel released those structures prematurely, any parallel read caused a use-after-free condition. An attacker who sends crafted TCP packets can leak kernel heap addresses, trigger out-of-bounds writes, or crash the machine. A failed race does not crash the host; it merely terminates the individual handshake. This behaviour allows an attacker to retry repeatedly over the network until the timing succeeds.
Upstream maintainer Eric Dumazet fixed the flaw by modifying how the kernel releases these option structures in an upstream Linux kernel commit. The four-line patch swaps immediate memory deallocations for read-copy-update calls, ensuring memory persists until all active read threads across every processor core finish their work.
AI-assisted exploitation and multipath TCP
The bug gained urgency when an Anthropic researcher reported that an artificial intelligence model produced a fully functional exploit against targeted configurations. Anthropic stated that Claude created an attack chain able to run “arbitrary commands as root, without any credentials or user interaction” on systems running multipath TCP listeners.
In Anthropic’s test environment, the exploit took 48 minutes to leak a kernel pointer over the network, followed by five minutes to secure an interactive root shell. Anthropic has withheld the exploit code under a standard coordinated disclosure window.
The requirement for a multipath TCP listener narrows the initial attack surface, but modern software stacks often enable the feature without explicit administrative configuration. In an advisory posted to the oss-security mailing list, the Go Security Team noted that “as of Go 1.24, Go programs enable MPTCP by default for listen sockets on Linux.”
Any server hosting network services compiled with modern Go toolchains on an MPTCP-capable kernel exposes that socket type automatically. The Go team cautioned that the underlying kernel bug is not restricted to multipath TCP, warning that standard TCP sockets could succumb to similar exploitation methods once researchers refine timing techniques.
Distribution patching status
Upstream kernel maintainers released patched point versions across all supported branches in mid-October 2026, including kernel lines 5.10, 5.15, 6.1, 6.6, 6.12, 6.18, 7.2, and 7.3. However, enterprise production servers rely on distribution packages with backported security updates, creating an uneven patching picture.
Debian has published updates for Debian 12 and Debian 13. Amazon Linux patched all standard and extras kernels across Amazon Linux 2 and Amazon Linux 2023 on 30 October. SUSE published resolved packages for SUSE Linux Enterprise Server 16 and several Long Term Service Pack Support releases, as documented in SUSE’s vulnerability tracker, while openSUSE Leap 16.0 also has an update.
By contrast, other widely deployed platforms remain exposed. Canonical lists the kernel package as vulnerable across all supported Ubuntu releases, with work-in-progress packages under test for releases 22.04 through 26.10. Red Hat upgraded its assessment to Critical with a CVSS 3.1 score of 9.8 in security advisory RHSB-2026-016, confirming that Red Hat Enterprise Linux versions 8, 9, and 10 remain affected while engineering teams expedite fixes.
AlmaLinux built replacement kernels for versions 8, 9, and 10 and placed them into public testing repositories ahead of Red Hat, pending verification before wide production deployment. Containers offer no insulation from the bug: because container runtimes share the host kernel, updating container base images does not protect workloads.
Interim firewall mitigations
Because the vulnerability requires the attacker to supply IPv4 options in the initial SYN segment, perimeter firewalls and local packet filters can stop exploit traffic before it reaches the kernel socket layer. Legitimate production TCP traffic rarely carries IPv4 options, making filtering practical.
Red Hat and community teams recommend implementing an nftables rule that drops incoming IPv4 TCP traffic where the IP header length exceeds five 32-bit words, which corresponds to 20 bytes. A header length above 20 bytes confirms the presence of IP options:
nft add table ip filternft add chain ip filter raw { type filter hook prerouting priority -300 \; }nft add rule ip filter raw ip protocol tcp ip hdrlength gt 5 drop
Placing the rule in the prerouting hook ensures that packets routed forward to local container namespaces are blocked alongside packets bound for the host. Administrators managing systems with traditional iptables can achieve identical filtering using the raw table and the u32 extension module.
Teams cannot apply these packet-filtering rules on networks that use Commercial IP Security Option labels, which travel inside IPv4 header fields. Filtered host boundaries also do not intercept bridged network traffic moving directly between two containers sharing the same internal virtual switch.
Disabling multipath listeners
Organisations unable to apply kernel updates or firewall filters immediately can disrupt known exploit mechanics by disabling multipath TCP support. Setting the kernel sysctl parameter net.mptcp.enabled to zero forces applications to revert to conventional TCP sockets.
Because Linux enforces sysctl network settings per namespace, administrators must apply this configuration to container runtimes alongside the parent operating system. Individual Go services can also disable MPTCP listeners by launching with the environment flag GODEBUG=multipathtcp=0, though security teams treat binary-level overrides as secondary precautions.
The public record on the National Vulnerability Database record continues to track vendor advisories as proof-of-concept scripts emerge across public software repositories. Systems requiring remediation must be rebooted after installing updated kernel packages to activate the memory fixes.
