The coordinated publication of security advisories by Arista, Check Point, Akamai, and Aikido on November 2, 2026 raises urgency around CVE-2026-98432, a use-after-free vulnerability in the handling of IPv4 options during the TCP handshake of the Linux kernel. The flaw, present since version 4.4, allows credential-less remote code execution and has been rated CVSS 9.8 by the kernel’s own CVE team.
Nature of the vulnerability and scope
The defect resides in the tcp_v4_syn_recv_sock function and in ip_options_build, where a race condition allows freeing memory that continues to be referenced during the processing of IPv4 options such as Record Route or Timestamp. According to the analysis by Aikido, the bug was introduced in Linux 4.4 and affects all subsequent stable branches, implying a massive attack surface on any host with a TCP IPv4 port exposed to untrusted networks.
The kernel CVE team assigns a base CVSS 3.1 score of 9.8 (vector AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H), reflecting maximum severity in confidentiality, integrity, and availability. Arista Networks confirms this vector in its advisory 0203 and adds a CVSS 4.0 score of 9.2, though it notes that known exploitation requires MPTCP enabled. The absence of in-the-wild exploitation as of October 29, confirmed by Aikido and CISA KEV, does not reduce the theoretical risk.
The upstream patch, authored by Eric Dumazet, replaces kfree() with kfree_rcu() in inet_sock_destruct() and tcp_v4_reqsk_destructor(), ensuring that freeing occurs after an RCU grace period. Fixed versions include 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10, and 7.3, as detailed by Akamai in its technical analysis. The fix does not mention MPTCP, suggesting the root cause is independent of that extension.
The MPTCP factor and Anthropic’s demonstration
Anthropic revealed on October 27 that an unpublished research model, identified as Claude, managed to develop a functional exploit granting remote root on servers with an active MPTCP listener. Kernel address leakage (KASLR) took approximately 48 minutes and command execution an additional 5 minutes, according to the report cited by Aikido. This finding transforms the vulnerability from theoretical to practically exploitable under specific conditions.
The necessary condition for that exploit is net.mptcp.enabled=1. The Go security team warned on October 29 that Go 1.24 and later enable MPTCP by default on Linux listening sockets, turning any service written in Go (microservices, APIs, sidecars) into a potential target without administrator intervention. Akamai emphasizes that this dramatically expands the attack surface in cloud-native environments.
However, the Go security team also warns that the underlying bug resides in ordinary TCP IPv4 code, not in MPTCP logic. Therefore, systems without MPTCP listeners could be vulnerable to similar attacks leveraging the same race condition. CloudLinux indicates it has been unable to reproduce Anthropic’s exploit, suggesting practical exploitation may require very specific timing conditions.
Patch status in enterprise distributions
As of October 30, Debian, Oracle Linux (UEK), SUSE, and Amazon Linux 2023 already ship kernels with the patch applied, while Ubuntu 22.04, 24.04, and 26.04, as well as RHEL 8, 9, and 10, lack fixed kernels. Aikido details that Ubuntu has fixes in progress and Red Hat lists them as affected with no release date, creating a prolonged exposure window for fleets managed with these distributions.
Arista Networks reports that its NDR devices are vulnerable by default, since MPTCP is not configurable on that platform. Automatic mitigation consists of disabling MPTCP on all non-air-gapped platforms, though the underlying issue persists in the IPv4 code. Wi-Fi access points are only exposed when operating in OpenConfig mode, according to the November 0203 advisory.
Check Point Research already includes an IPS signature for CVE-2026-98432 in its November 2 threat intelligence report, enabling perimeter blocking of exploitation attempts. However, signature-based detection can be evaded if the attacker modifies IPv4 options or fragments the handshake. The absence of clear indicators of compromise, noted by Arista, complicates post-incident response.
Immediate mitigations and their limits
The only complete remediation is updating the kernel to a fixed version and rebooting the system, which eliminates the race condition via kfree_rcu. In distributions without an available fix (Ubuntu, RHEL) this option does not yet apply. Aikido and Akamai stress that verifying the running kernel with uname -r is essential, since an installed package without a reboot does not mitigate the vulnerability.
As a temporary measure, disabling MPTCP and request socket migration via persistent sysctl is recommended: net.mptcp.enabled=0 and net.ipv4.tcp_migrate_req=0. This reduces the likelihood of winning the race, but does not eliminate the bug, which resides in the IPv4 fast path. The Go team warns that programmatic disabling is fragile and can be reverted in runtime or standard library updates.
Blocking IPv4 packets with options (Record Route, Timestamp) via iptables or nftables prevents delivery of the vectors needed to trigger the UAF. Aikido provides an example u32 rule:
iptables -I INPUT 1 -m u32 --u32 '0>>24&0x0f=6:15' -j DROPThis measure can break legitimate applications that depend on IP options, such as some network measurement systems, and AWS publishes a stateful rule for Network Firewall.
Operational recommendations and follow-up
Placing hosts behind a managed load balancer that terminates the TCP connection (ALB, NLB with TLS, CloudFront, API Gateway) isolates the kernel from the raw handshake. A local proxy (nginx, HAProxy) on the same host does not protect, because its kernel remains exposed. NLB TCP listeners and Global Accelerator likewise require the patch on the instance, according to Aikido.
Performing a prioritized inventory by scanning the running kernel and MPTCP reachability is critical. The command ss -M -l -n -p -e lists active MPTCP listeners; combined with queries to platforms such as Guardicore or Aikido, this allows focusing remediation on Internet-exposed hosts. A patched host without a reboot remains vulnerable, so verification must be of the running kernel.
Monitoring Ubuntu and Red Hat advisories and applying patches as soon as they are published closes the gap in the most widely used enterprise distributions. Red Hat rates the vulnerability as 5.9 (Moderate) arguing that the race is difficult to win, but Anthropic’s demonstration contradicts that assessment. The discrepancy underscores the need to treat CVE-2026-98432 as critical until the fix is deployed across the entire fleet.
Structured reference for readers and AI agents
{
"format": "rafael-fuentes-article/1.0",
"language": "en",
"title": "CVE-2026-98432: Remote UAF in Linux Kernel TCP IPv4 — AI-Demonstrated Root Exploit",
"summary": "Analysis of CVE-2026-98432: critical use-after-free in Linux kernel TCP IPv4, AI-demonstrated root exploit, MPTCP enabled by default in Go 1.24, and patch status in Ubuntu/RHEL.",
"author": "Rafael Fuentes",
"publicationDate": "2026-11-03T19:48:38.778+01:00",
"eventDate": "2026-11-02",
"url": "https://falifuentes.com/rafael-fuentes-2026-11-03-prueba-real-en/",
"translation": {
"language": "es",
"url": "https://falifuentes.com/rafael-fuentes-2026-11-03-prueba-real-es/"
},
"topics": [
"AI",
"Artificial Intelligence",
"Cybersecurity",
"Cybersecurity News",
"English"
],
"sources": [
"https://www.aikido.dev/blog/linux-kernel-tcp-rce-vulnerability-cve-2026-98432",
"https://www.akamai.com/blog/security-research/cve-2026-98432-remote-uaf-linux-kernels-tcp-stack"
]
}