CVE-2026-98432: the Linux kernel TCP flaw, who is at risk and how to close it - init.d - Linux systems, DevOps and PrestaShop development

# CVE-2026-98432: the Linux kernel TCP flaw, who is at risk and how to close it

Alessandro Corbelli~14 min
Table of Contents

On 15 October 2026 the Linux kernel CVE team published CVE-2026-98432, a flaw in the code that accepts incoming TCP connections over IPv4, and the CNA's entry gives it a score of 9.8 out of 10. On 27 October Anthropic stated that one of its research models had written an exploit that, by sending packets to a Linux server with an MPTCP listener, runs commands as root with no credentials. Hence the headlines claiming that every Linux server with an open port is at risk. That is only partly true, and the true part is enough to act on.

The faulty code lives in the kernel, the core of the operating system, and has been there since kernel 4.4, released in January 2016: it concerns almost every Linux server in production, whatever software runs on top. Below: what goes wrong, who is exposed and under which conditions, the patch status verified on 6 November 2026 at 12:12 UTC, how to verify and update, and in short the firewall stopgap for those who cannot reboot right away, which has an article of its own. Where sources disagree, we say so.

What goes wrong in the kernel, in plain words and then in detail

Every TCP connection starts with a handshake: the client sends a SYN packet ("I would like to connect"), the server replies with a SYN-ACK, the client confirms. Meanwhile the server keeps a note about the request. If the SYN carries IP options, an optional field of the IPv4 header designed decades ago for network diagnostics (for example "record the route"), the server stores them in that note to reuse them in the reply.

The mistake is in the handover: when the connection completes, the kernel moves the options from the note to the real connection without making a copy. If at that very moment a retransmitted SYN arrives and another processor is handling it, that processor reads the options the first one has already thrown away. That is a use-after-free, using memory after releasing it: like a clerk reading a sheet of paper while a colleague has already shredded it and the paper has been reassigned to another file. He reads data that no longer belongs there and, in some cases, writes on it.

The technical detail, from Eric Dumazet's fix commit: tcp_v4_syn_recv_sock() transfers ireq->ireq_opt to the child socket without duplicating it; a retransmitted SYN processed concurrently builds the SYN-ACK reading those options under rcu_read_lock() only, while inet_sock_destruct() frees them with a plain kfree(), and ip_options_build() reads freed memory. A similar variant, in tcp_v4_reqsk_destructor(), concerns request migration between listening sockets, active only with net.ipv4.tcp_migrate_req=1 or a BPF program. The fix is two lines: kfree() becomes kfree_rcu() in net/ipv4/af_inet.c and net/ipv4/tcp_ipv4.c, so the memory is released only once every reader is done. The defect dates back to commit 079096f103fa, from October 2015, which shipped in Linux 4.4. IPv6 is not affected by the first issue; for the migration variant alone the commit announces a separate fix.

What an attacker can achieve, in order of difficulty: crash the machine (denial of service), read fragments of kernel memory, corrupt it (that is Red Hat's assessment) and, in the case described by Anthropic, run commands as root: 48 minutes to leak a kernel pointer, the step that defeats kernel address randomisation (KASLR), and another five to reach command execution. The CVE record justifies the low attack complexity with the fact that the attacker can create option-bearing requests and repeat the race between processors as many times as needed.

Who is at risk: affected kernels and distribution status on 6 November 2026

According to the CVE record, every kernel from 4.4 onwards without the fix is vulnerable; the fixed stable branches are 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and the upcoming 7.3. Distributions carry the fix onto their own version: the advisory is what counts, not the number, as with the distributor's backported patches.

DistributionFixed kernel (or later)Advisory and date
Debian 13 "trixie"linux 6.12.113-1DSA-6568-1, 16 October
Debian 12 "bookworm"linux 6.1.190-1; linux-6.12 6.12.113-1~deb12u1DLA-4841-1 and DLA-4850-1 (LTS)
Ubuntu 24.04 LTSlinux 6.8.0-149.149in noble-updates since 6 November, 01:14 UTC; no USN at the time of checking
Ubuntu 22.04 LTSlinux 5.15.0-203.213in jammy-updates since 5 November, 23:22 UTC; no USN at the time of checking
Ubuntu 26.04 LTSlinux 7.0.0-43.43in resolute-updates since 6 November, 01:14 UTC; no USN at the time of checking
Ubuntu 20.04 LTSnonefix pending according to Canonical; HWE kernel 5.15.0-203.213~20.04.1 work in progress
RHEL 8.10kernel 4.18.0-553.182.1.el8_10RHSA-2026:93414 , 1 November
RHEL 9.8kernel 5.14.0-687.64.1.el9_8RHSA-2026:93412 , 1 November
RHEL 10.2kernel 6.12.0-211.72.1.el10_2RHSA-2026:93413 , 1 November
AlmaLinux 8, 9, 10same versions as RHELproduction repositories since 2 November
Rocky Linux 8, 9, 10same versions as RHELRLSA-2026:93414 , 93412 and 93413, 3 November
CloudLinux 7h and 84.18.0-553.180.1.lve.1since 30 October; CL9 (5.14.0-687.64.1.el9_8) and CL10 (6.12.0-211.73.1.el10_2) in the stable channel since 3 November
Oracle Linux 8, 9, 10see Oracle's CVE pageELSA-2026-93414, 93412 and 93413, 2 November; UEK kernels since 16 October
Amazon Linux 2 and 2023see Amazon's CVE pageALAS advisories from 24 to 30 October
SUSE SLES 15 SP4-SP7 and 16.0see SUSE's CVE pageSUSE-SU advisories from 29 October to 5 November

Notes. On Ubuntu the picture changed in the last few hours: Canonical's blog post of 2 November talked about -proposed and the CVE page updated on the 5th still marks the three versions as work in progress, but the Launchpad API shows them published to -updates between 23:22 UTC on 5 November and 01:14 on 6 November, in an Ubuntu 24.04 container apt-cache policy offers 6.8.0-149.149 from noble-updates and that kernel's changelog contains the fix commit. At the time of checking there is no USN and -security still holds the previous kernel: the default unattended-upgrades configuration on Ubuntu (/etc/apt/apt.conf.d/50unattended-upgrades, read on 24.04) does not include -updates, so anyone relying on automatic updates does not get this kernel until the USN is out and has to install it by hand. For the HWE kernels (6.8 on 22.04, 5.15 on 20.04) Launchpad still shows the previous versions in -updates (6.8.0-149.149~22.04.1 has been in jammy-proposed since 29 October) and the CVE page marks the fixed ones as work in progress. RHEL minor releases under extended support have their own advisories on Red Hat's CVE page. Debian 11 has been out of LTS since 31 August 2026 and does not appear in Debian's tracker: we do not know its status. RHEL 7 (Red Hat marks it as not affected) and CloudLinux 7 (not affected according to CloudLinux) do not need the fix.

Is an open TCP port enough? The real conditions

Two versions circulate on blogs, "it needs MPTCP" and "any TCP port will do". The primary sources settle it.

  • Any service listening on IPv4 TCP is a target. The commit never mentions MPTCP; Red Hat's bulletin states that RHEL systems with MPTCP disabled, as it is by default, remain affected; AlmaLinux writes that "The bug itself is in the regular IPv4 TCP code, so systems without MPTCP still need the fix", and the Go security team writes that "It is possible that systems with no MPTCP listeners are vulnerable to similar attacks". Anthropic's exploit uses an MPTCP listener (Multipath TCP, the extension that lets one connection travel over several paths): it is the demonstrated route to code execution, not the only route to the defect. On a plain TCP listener we know of no public demonstration of code execution; Red Hat justifies the Critical rating with memory corruption and crashes reachable over ordinary IPv4 TCP traffic.
  • The SYN must carry IPv4 options. Without options the structure that gets freed too early does not exist: every distributor's mitigation relies on this.
  • IPv4 only. A service reachable only over IPv6 is not involved.
  • The "migration" variant is off by default: net.ipv4.tcp_migrate_req is 0 unless you changed it.
  • Internal networks included. Anyone who can send a SYN to the service can try: an infected laptop on the LAN is as good as the internet.
  • Containers and virtual machines. A container uses the host's kernel: a port published by Docker on a vulnerable host is vulnerable, whatever runs inside. A VM depends on its own kernel, not the hypervisor's.
  • Routers, NAT and network firewalls. RFC 7126 speaks of a "widespread practice of blocking IPv4 optioned packets", but widespread does not mean guaranteed: you do not know what your provider does, and a Linux router doing NAT has listening services of its own. Who actually uses IP options, and how rare they are, is in the article on the stopgap.
  • Cloud security groups. An AWS security group rule specifies protocol, ports, ICMP type and source: there is no field for IP options, so it does not replace the filter described below.

What we know and what we don't

Certain (primary sources)Stated limits
Fix by Eric Dumazet on a report by Xinyang Ge (anthropic.com address in the commit), written on 1 October and applied on the 5th, available in Linux 7.3-rc7Anthropic did not publish code, kernel version or exploit configuration with its article; we know of no independent reproduction
CVSS 9.8 from the kernel CVE team and Red Hat (Critical); 8.1 from SUSE and Amazon (Important) and from Oracle (preliminary)Which of the two numbers describes the risk better depends on how hard the race between processors is to win, and Red Hat itself writes that reliable code execution requires precise control of memory
Anthropic claims a root exploit against a server with an MPTCP listenerExploitation in the wild: none of the primary sources we read reports it
Not in the CISA KEV catalog (version 2026.11.04, checked on 6 November at 12:12 UTC)The catalog only records confirmed exploitation: absence is not a guarantee
The firewall rule looks only at IP options: in our tests HTTP traffic, even inside a WireGuard tunnel, never touches it, and the kernel already drops source routingHow many packets with IP options actually reach a production server: the last measurement we know of is from 2005, and we made none on your server

On the numbers: 9.8 and 8.1 differ by a single entry of the CVSS vector, attack complexity: AC:L for the kernel team and Red Hat, AC:H for SUSE, Oracle and Amazon. Red Hat justifies the Critical rating with the absence of prerequisites beyond IPv4 TCP traffic and the risk of severe instability, while conceding that reliable code execution requires precise control of memory.

If you cannot update right away: the firewall stopgap in short

Red Hat in its bulletin and AlmaLinux on its blog publish the same idea: drop IPv4 TCP packets with IP options, that is with a header longer than 20 bytes (ip hdrlength > 5, in 32-bit words). Neither presents it as a fix: Red Hat speaks of reducing exposure, AlmaLinux writes "Neither step is a fix", and the Ubuntu Security Team that "security updates are the only appropriate mitigation". AlmaLinux's variant hooks the prerouting point, so it also covers ports published by Docker, which are diverted in the nat table before they reach input; in exchange it also drops forwarded traffic, if the machine acts as a router. In the file /etc/cve-2026-98432.nft:

table ip cve_2026_98432 {
chain prerouting {
type filter hook prerouting priority -300; policy accept;
meta l4proto tcp ip hdrlength > 5 counter drop
}
}
Terminal window
sudo nft -c -f /etc/cve-2026-98432.nft
sudo nft -f /etc/cve-2026-98432.nft
sudo nft list table ip cve_2026_98432
sudo nft delete table ip cve_2026_98432

The four commands check, load, show and remove the rule; the counter tells you how many packets it dropped. We tested them in a Debian 13 container and on a Docker host with a published port: HTTP traffic never touches the counter, a SYN with the "record route" option is counted and, with drop, dropped. The table does not survive a reboot and, in our test on a Docker host, does not see traffic between two containers on the same bridge: AlmaLinux warns that such traffic may bypass it. The iptables and ufw variants, persistence, what it can break, how to introduce it in observation mode first and the other mitigations (MPTCP off, livepatch, init_on_free) are in the article on the firewall stopgap: here it is enough to say that it is a bridge to the maintenance window, not a way to skip it.

How to check whether you are vulnerable and how to update

The running kernel is shown by uname -r; the real check is the package changelog, because distributions do not change the upstream version number, as with PHP2Shell. Before touching production, a verified backup.

On Debian the kernel changelog is changelog.gz in the package's documentation folder (verified on Debian 13 with package 6.12.115-2, where the line sits in the 6.12.113-1 block):

Terminal window
uname -r
sudo apt update && sudo apt install linux-image-amd64
zcat /usr/share/doc/linux-image-$(uname -r)/changelog.gz | grep -m1 CVE-2026-98432

On Ubuntu the changelog lists the fix by the commit title, on 24.04 in the "Noble update: upstream stable patchset 2026-10-21" block, not by CVE number: searching for CVE-2026-98432 finds nothing. First check that the repository offers the fixed version:

Terminal window
apt-cache policy linux-image-generic
sudo apt update && sudo apt install linux-generic
apt-get changelog linux-modules-$(uname -r) | grep -m1 'inet_opt and ireq_opt'

On AlmaLinux 9 the package changelog cites the CVE (verified with kernel-core 5.14.0-687.66.1.el9_8; we did not check RHEL and Rocky), and the line reads "ipv4: free inet_opt and ireq_opt after an RCU grace period (CKI Backport Bot) [RHEL-295635] {CVE-2026-98432}":

Terminal window
sudo dnf clean metadata && sudo dnf upgrade kernel
rpm -q --changelog kernel-core-$(uname -r) | grep -m1 CVE-2026-98432

Then reboot: the new kernel is not active until the machine restarts, and uname -r after the reboot must show the fixed version (dnf needs-restarting -r, from the dnf-plugins-core package, and Ubuntu's /var/run/reboot-required file tell you whether it is needed: both tried in the containers). On Ubuntu 20.04 the fix is pending: standard support ended in May 2025 and security updates arrive through Ubuntu Pro (ESM), free for personal use on up to five physical machines. It is also the moment to count your exposed ports: every reachable service is a target, and going through a bastion host or a VPN takes quite a few off the list. If you would rather have someone else do it, the server security check covers exactly exposed ports, firewall and missing updates, on quote once we have seen the scope.

In summary

CVE-2026-98432 is a use-after-free in the Linux kernel's handling of IPv4 TCP connections, present since 2016 and fixed with two lines. It affects any service listening on IPv4 that receives a SYN with IP options, not just MPTCP, which is the route Anthropic demonstrated to reach root. As of 6 November, fixes exist for Debian, RHEL 8, 9 and 10 with AlmaLinux, Rocky and CloudLinux, SUSE, Oracle and Amazon Linux; for Ubuntu 22.04, 24.04 and 26.04 the fixed kernel is in -updates but without a USN, so default automatic updates do not install it; they are still missing for Ubuntu 20.04, where they will arrive through Pro. It is not in the CISA KEV catalog and Anthropic's exploit was not published with its article, but the race can be retried. The nftables rule is a stopgap to reach the maintenance window: it drops only packets with IP options, which ordinary traffic does not carry; the new kernel and the reboot remain the only fix, and the proof is in the changelog, not in uname -r. If a NIS2 client asks how you handle patches, this case is a good example of a vulnerability plan to tell with dates and advisories.

Sources

Tux versione Gandalf, mascotte del blog init.d

init.d is the team led by Alessandro Corbelli, a Linux systems administrator and backend developer with over 15 years of experience. He designs and runs cloud infrastructure (Google Cloud, AWS, Azure), server farms and high-availability architectures, and builds custom software in Laravel/PHP and Vue - from the Take2Me food delivery platform to our clients' management tools. In this changelog we share technical notes on Linux, system administration, development, DevOps and e-commerce.


More Posts
Book a 30-minute call
Open in a new tab