No attacks reportedFix released · 11 Oct 2026
Linux kernel · IPv4 TCP handshake
CVE-2026-98432 Severity · Critical
Unauthenticated remote use-after-free · CVSS 9.8
Release lineFixed in
7.37.3
7.27.2.10
6.186.18.56
6.126.12.113
6.66.6.159
6.16.1.190
5.155.15.223
5.105.10.272
5.4Vendor fixes only: Amazon, Oracle UEK6
4.19No upstream fix
4.14Vendor fix only: Amazon Linux 2
4.4 and 4.9No upstream fix

CVE alert · Linux

CVE-2026-98432: a race in Linux's IPv4 TCP handshake lets remote attackers corrupt kernel memory

A race in Linux's IPv4 TCP handshake code, rated 9.8 Critical by the kernel's CVE team, lets a remote attacker with no login corrupt kernel memory. Upstream, Debian, Oracle and Amazon Linux ship fixes; Red Hat and Ubuntu have none as of 1 November 2026.

Check your systems with your AI agent

Paste this prompt into an AI agent that can run commands on your machines. It only reads versions and settings, asks before using administrator rights, and ends with a report.

Check whether my systems are affected by CVE-2026-98432 in Linux kernel, then write me a report.

Read the alert first: https://securityaudit.ae/cve/cve-2026-98432-linux-kernel/

What the alert says (facts checked 1 November 2026):
- Affected: Linux 4.4 and later, before 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3 (CVE record). Kernels before 4.4 are not affected. Distribution kernels have their own version numbers: check the vendor's fixed package.
- Fixed in: Upstream: 6.18.56 and 7.2.10 (11 October 2026), 5.10.272, 5.15.223, 6.1.190, 6.6.159 and 6.12.113 (15 October 2026), 7.3 (18 October 2026). Distribution fixes as of 1 November 2026: Debian, Oracle UEK, Amazon Linux and part of SUSE; see Who is affected.
- No fix yet: As of 1 November 2026: Red Hat Enterprise Linux 8, 9 and 10 and OpenShift; Ubuntu (for the main linux package, fixes in progress for 22.04 to 26.10 and none for 16.04 to 20.04; other kernel packages differ, see Who is affected); Rocky Linux; AlmaLinux outside its testing repository; SUSE Linux Enterprise Server 15 SP7 and 12 SP5 and SUSE Linux Micro 6.0 and 6.1; Oracle's Red Hat Compatible Kernel (none listed). Upstream gives none for end-of-life lines such as 5.4 and 4.19.
- Exploited: No attacks reported as of 1 November 2026. CISA's Known Exploited Vulnerabilities (KEV) catalog, version 2026.10.30, does not list the CVE. Anthropic built a working exploit; its write-up links no code for it. Seven public GitHub repositories describe a proof of concept or a trigger; none claims remote code execution.

Rules:
- Read only. Do not install, update, restart, delete or change anything, and never run exploit or proof-of-concept code.
- Run only commands that read versions or settings. Ask me before any command that needs sudo or administrator rights.
- If you do not know which machines, containers, cloud accounts or sites are in scope, ask me for the list before you start.
- If a machine cannot be reached or a command fails, say so. Do not guess.

Steps:
1. Find where Linux kernel runs: servers, containers and their hosts, cloud services, sites.
2. On each one, read the installed version and the settings the alert names. The alert gives these commands:

    uname -r

    # Debian and Ubuntu
    dpkg -l 'linux-image*'
    # RHEL, AlmaLinux, Rocky Linux, Oracle Linux (RHCK), Amazon Linux
    rpm -q kernel
    # Amazon Linux 2023 with the 6.12 or 6.18 kernel
    rpm -q kernel6.12 kernel6.18
    # Oracle Linux with UEK
    rpm -q kernel-uek
    # dnf systems, after an update: exit code 1 means a reboot is needed
    sudo dnf needs-restarting -r

    sysctl net.mptcp.enabled
    sudo ss -lnMp

    go version -m /path/to/binary
    go list -f '{{.DefaultGODEBUG}}' ./path/to/main/package

    sysctl net.ipv4.tcp_migrate_req

    sudo nft list table inet cve_2026_98432   # Red Hat's rule
    sudo nft list table ip cve_2026_98432     # AlmaLinux's rule

3. Compare each version with the affected and fixed versions above. When a vendor or a distribution ships its own fixed package, the alert names it.
4. Report: one line per machine or service with the version found, Affected, Not affected or Unknown, and the evidence (the command and its output). Then the actions to take, in order, from the alert's "How to fix" section, and what you could not check.

Overview

CVE-2026-98432 is a use-after-free in the Linux code that accepts incoming IPv4 TCP connections. Xinyang Ge of Anthropic reported it, and the first stable releases with Eric Dumazet's fix came out on 11 October 2026. The Linux kernel CNA (CVE Numbering Authority, here the kernel's own CVE team) published the CVE on 15 October 2026 and added its CVSS score nine days later.

On 27 October 2026, Anthropic's Frontier Red Team described an exploit that lets a remote attacker send packets "to Linux servers with a multipath TCP listener to run arbitrary commands as root, without any credentials or user interaction". Multipath TCP (MPTCP) lets one connection use several network paths at once. The Go Security Team, Red Hat (RHSB-2026-016) and AWS (2026-160-AWS) published advice three and four days later.

As of 1 November 2026, every maintained upstream line has the fix, and so do Debian, Oracle's Unbreakable Enterprise Kernel (UEK), Amazon Linux and part of SUSE. Red Hat, Ubuntu and Rocky Linux have released none, and AlmaLinux's are still in testing. We tested the Red Hat and AlmaLinux firewall rules and the check commands in Docker containers.

CVE-2026-98432 · Linux kernelCritical · CVSS 9.8

Affected
Linux 4.4 and later without the fix, on any system that accepts IPv4 TCP connections; containers run on their host's kernel. As of 1 November 2026, Red Hat Enterprise Linux 8 to 10 and Ubuntu 16.04 to 26.10 have no fixed kernel.
Fix
Install a fixed kernel and reboot. Upstream: 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56 or 7.3, or a later release of the same line; move off 7.2, which reached end of life. Debian, Oracle UEK, Amazon Linux and part of SUSE ship fixed packages as of 1 November 2026.
Can't update yet
Red Hat, AlmaLinux and AWS give firewall rules that drop IPv4 packets carrying IP options; Red Hat says its rule "cannot fully prevent" the attack. AlmaLinux and AWS also suggest turning MPTCP off. See If you can't update yet.

Why it's urgent: The kernel CNA says the bug is reached "during the TCP handshake, before application authentication", on any listening IPv4 TCP port. Anthropic's Frontier Red Team built a working exploit that ran commands as root on servers with an MPTCP listener. As of 1 November 2026, Red Hat rates the flaw Critical, as an unauthenticated remote attacker can "cause kernel memory corruption or an immediate system crash", and neither it nor Ubuntu has shipped a fixed kernel.

Key facts

FactDetail
ProductLinux kernel (operating system kernel)
Severity9.8 Critical (CVSS 3.1 AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H, the Linux kernel CNA, added to the record on 24 October 2026). Other ratings as of 1 November 2026: Red Hat, Critical (9.8, draft); SUSE, important (8.1 in CVSS 3.1 with AC:H, high attack complexity, and 9.2 in CVSS 4.0); Oracle Linux (preliminary) and Amazon Linux, Important (8.1, AC:H); Ubuntu, High priority, with a CVSS 4.0 score of 8.3 on its page; NVD, no score of its own.
WeaknessUse-after-free in ip_options_build() after a race between two CPUs (CVE record). The record names no CWE; Red Hat assigns CWE-825 (Expired Pointer Dereference).
ImpactThe kernel CNA says the stale options can "disclose adjacent kernel heap contents directly to the peer", allow "out-of-bounds kernel writes and potential control-flow corruption", and make the kernel panic. Red Hat: "kernel memory corruption or an immediate system crash". Anthropic's exploit ran shell commands as root on servers with an MPTCP listener.
Who can attackAnyone who can send packets to a listening IPv4 TCP port, with no login and no user action. The connection request (SYN) must carry IP options, and the attacker must win a race that Red Hat calls "timing-sensitive". Anthropic's root exploit needed an MPTCP listener.
AffectedLinux 4.4 and later, before 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3 (CVE record). Kernels before 4.4 are not affected. Distribution kernels have their own version numbers: check the vendor's fixed package.
Fixed inUpstream: 6.18.56 and 7.2.10 (11 October 2026), 5.10.272, 5.15.223, 6.1.190, 6.6.159 and 6.12.113 (15 October 2026), 7.3 (18 October 2026). Distribution fixes as of 1 November 2026: Debian, Oracle UEK, Amazon Linux and part of SUSE; see Who is affected.
No fix yetAs of 1 November 2026: Red Hat Enterprise Linux 8, 9 and 10 and OpenShift; Ubuntu (for the main linux package, fixes in progress for 22.04 to 26.10 and none for 16.04 to 20.04; other kernel packages differ, see Who is affected); Rocky Linux; AlmaLinux outside its testing repository; SUSE Linux Enterprise Server 15 SP7 and 12 SP5 and SUSE Linux Micro 6.0 and 6.1; Oracle's Red Hat Compatible Kernel (none listed). Upstream gives none for end-of-life lines such as 5.4 and 4.19.
ExploitedNo attacks reported as of 1 November 2026. CISA's Known Exploited Vulnerabilities (KEV) catalog, version 2026.10.30, does not list the CVE. Anthropic built a working exploit; its write-up links no code for it. Seven public GitHub repositories describe a proof of concept or a trigger; none claims remote code execution.
Proof of conceptAnthropic's Frontier Red Team built an exploit that ran commands as root on servers with an MPTCP listener, described in its write-up of 27 October 2026, which links no code for it. Seven public GitHub repositories describe a proof of concept or a trigger, going by their descriptions; we did not open or run their code.
CVE recordTitle "ipv4: free inet_opt and ireq_opt after an RCU grace period", from the Linux kernel CNA, published on 15 October 2026. As of 1 November 2026, NVD lists it as Received, with no score or analysis of its own.
CreditReported by Xinyang Ge of Anthropic. Fixed by Eric Dumazet, reviewed by Kuniyuki Iwashima and applied by Jakub Kicinski. Greg Kroah-Hartman handled the stable backports and the CVE announcement.
CVE published15 October 2026

Who is affected

Per the CVE record, kernels from 4.4 on are affected unless they include the fix, which arrived in 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3. Older kernels are not affected. Upstream fixes stop at the 5.10 line, so end-of-life lines below it, such as 5.4, 4.19 and 4.14, get no fix from the kernel developers; Amazon fixes its own 4.14 and 5.4 kernels, and Oracle its 5.4-based UEK6. Distribution kernels have their own version numbers, so compare yours with the vendor's fixed package below.

Release lineStatus on 1 Nov 2026Fixed inReleased
7.3Stable7.318 Oct 2026
7.2End of life7.2.10 (Ended 1 November 2026: move to 7.3)11 Oct 2026
6.18Longterm6.18.5611 Oct 2026
6.12Longterm6.12.11315 Oct 2026
6.6Longterm6.6.15915 Oct 2026
6.1Longterm6.1.19015 Oct 2026
5.15Longterm5.15.22315 Oct 2026
5.10Longterm5.10.27215 Oct 2026
5.4End of lifeVendor fixes only: Amazon, Oracle UEK6None
4.19End of lifeNo upstream fixNone
4.14End of lifeVendor fix only: Amazon Linux 2None
4.4 and 4.9End of lifeNo upstream fixNone

Debian

As of 1 November 2026, Debian has fixed trixie and bookworm through its security suites:

  • Debian 13 (trixie): linux 6.12.113-1, DSA-6568-1 of 16 October 2026.
  • Debian 12 (bookworm): linux 6.1.190-1, DLA-4841-1 of 17 October 2026, and linux-6.12 6.12.113-1~deb12u1, DLA-4850-1 of 22 October 2026.

Testing (forky) and unstable (sid) are fixed too. The security tracker still lists the base suites' 6.12.107-1 and 6.1.176-1 as vulnerable.

Ubuntu

As of 1 November 2026, Canonical has released no fixed kernel and no security notice (USN) for this CVE. Its CVE page shows fixes in progress for the main linux package: 7.3.0-10.10 on 26.10, 7.0.0-43.43 on 26.04 LTS, 6.8.0-149.149 on 24.04 LTS and 5.15.0-203.213 on 22.04 LTS. On 20.04, 18.04 and 16.04 the package is vulnerable, with no fix in progress. Cloud kernels are also being fixed: linux-aws, linux-azure and linux-gcp on 22.04 to 26.10, and their 5.15-based versions on 20.04. Some kernels have no fix in progress, such as the 24.04 HWE kernels linux-hwe-6.17 and linux-hwe-7.0.

Red Hat, AlmaLinux and Rocky Linux

Red Hat lists RHEL (Red Hat Enterprise Linux) 8, 9 and 10, RHEL for NVIDIA 26 and the RHEL CoreOS images of OpenShift Container Platform 4 and 5 as affected. As of 1 November 2026 it has published no fixed kernel and no security advisory (RHSA). Its bulletin says "Red Hat is expediting the release of fixes" and that any product relying on that kernel, RHEL CoreOS included, "is also potentially impacted". Its CVE page lists RHEL 6 and 7 as not affected; its VEX data (machine-readable status) gives the reason: the vulnerable code is not present.

AlmaLinux put fixed kernels in its testing repository on 30 October 2026, ahead of any RHEL fix: kernel-4.18.0-553.180.2.el8_10 for release 8, kernel-5.14.0-687.61.2.el9_8 for 9 and kernel-6.12.0-211.69.2.el10_2 for 10. As of 1 November 2026 its errata list no advisory for the CVE, and Rocky Linux has published none.

SUSE

SUSE rates the flaw important, and its CVE page gives the overall state as "Pending" as of 1 November 2026. SUSE's advisories date the SLES (SUSE Linux Enterprise Server) 16.0 and SUSE Linux Micro 6.2 kernels 24 October and the SLES 15 SP4 to SP6 kernels 29 October; its CVE page lists all of them as published on 29 October. Fixed kernels, at these versions or later, came out from 24 October 2026:

  • SLES 16.0, SUSE Linux Micro 6.2 and openSUSE Leap 16.0: kernel-default 6.12.0-160000.40.1.
  • SLES 15 SP6-LTSS: 6.4.0-150600.23.143.1 (SUSE-SU-2026:4934-1).
  • SLES 15 SP5-LTSS: 5.14.21-150500.55.195.1 (SUSE-SU-2026:4936-1).
  • SLES 15 SP4-LTSS: 5.14.21-150400.24.246.1 (SUSE-SU-2026:4937-1).

SLE Micro 5.4 and 5.5 and SLES for SAP 15 SP4 to SP6 are fixed too. SLES 15 SP7 is in progress, with 6.4.0-150700.53.90.1 in quality assurance. SLES 12 SP5 and SUSE Linux Micro 6.0 and 6.1 are listed as affected. SUSE's CVE page also lists openSUSE Leap 15.6 as affected, but SUSE-SU-2026:4934-1 ships it kernel-default 6.4.0-150600.23.143.1 (zypper in -t patch SUSE-2026-4934). Most livepatches (fixes applied to the running kernel, with no reboot) are not out yet.

Oracle Linux and Amazon Linux

Oracle fixed UEK on 16 October 2026, as its CVE page and advisories show: UEK6 kernel-uek-5.4.17-2136.360.3.1 for Oracle Linux 7 and 8 (ELSA-2026-500397), UEK7 kernel-uek-5.15.0-325.221.6.4 for 8 and 9 (ELSA-2026-500398), and UEK8 kernel-uek-6.12.0-207.111.5.4 for 9 and 10 (ELSA-2026-500399). As of 1 November 2026, the page lists no fix and no status for the Red Hat Compatible Kernel (RHCK).

Amazon Linux released fixes for all its listed kernels between 24 and 30 October 2026. Some carry the fix on a base below the upstream fixed release (5.10.271, for example), so compare your kernel with the package versions of its advisories:

Cloud services and containers

Containers run on their host's kernel, so patch and reboot each container host and Kubernetes node. The AWS bulletin, last updated on 1 November 2026 at 04:15 Dubai time, says:

  • EC2 and Lightsail: update the kernel of your instances and reboot them.
  • EKS: the EKS-optimized Amazon Linux 2023 AMIs (machine images) of release v20261031 fix the issue. Auto Mode AMIs were expected by 1 November 2026, with no action needed.
  • ECS on EC2: update and reboot the container instances, or replace them from ECS-optimized AMI release 20261031. Fargate tasks launched after 17:00 Dubai time on 1 November 2026 use the patched platform.
  • Bottlerocket (updated AMIs) and Redshift (OS patch): expected from 2 November 2026.
  • RDS and Aurora: make sure the database ports are not exposed to the internet. AWS is deploying a mitigation; the database OS patch, which you apply, is due on 3 November 2026, the day the Storage Gateway update is also expected to start.
  • Lambda: not affected, as functions expose no public IPv4 TCP port.
  • WAF, ALB, CLB and NLB TLS listeners, CloudFront and API Gateway: AWS is updating the infrastructure, and "no immediate customer action is required".

How the flaw works

A TCP connection opens with a handshake: the client sends a SYN (connection request), the server answers with a SYN-ACK, and the client's ACK completes it. A SYN may carry IP options (optional fields of the IPv4 header); the kernel CNA names Record Route and Timestamp. The kernel keeps them with the pending request for its SYN-ACK.

The fix's commit message describes the race. When the handshake completes, tcp_v4_syn_recv_sock() hands the stored options to the new connection's socket (the child socket) without copying them. Meanwhile another CPU can be handling a retransmitted copy of the SYN and building a SYN-ACK from the same options, protected only by RCU (read-copy-update, a kernel scheme that lets readers work without locks while frees wait for them). If the child socket is closed first, inet_sock_destruct() frees the options with a plain kfree(), and ip_options_build() keeps using freed memory.

Per the CNA's scoring notes, copying the stale options into the SYN-ACK can "disclose adjacent kernel heap contents directly to the peer". Stale lengths and offsets can steer writes, "permitting out-of-bounds kernel writes and potential control-flow corruption", and a stale length can make the kernel panic. The CNA rates the attack complexity low, since "Both racing sides are attacker-induced"; SUSE, Oracle and Amazon Linux rate it high.

Anthropic's exploit works in two stages. The first uses the race to make the kernel send back extra bytes until one reply leaks a kernel pointer, which defeats KASLR (kernel address space layout randomization, which loads the kernel at a random address). The second overwrites a callback pointer with the address of a payload staged in kernel memory, which ends by running a shell command as root. Anthropic notes that "a lost race only fails the handshake without crashing the whole system", so the attacker can retry. Its proof of concept took 48 minutes to leak a pointer and five more minutes to run the command.

That exploit needs an MPTCP listener. The Go Security Team says "the underlying bug is not specific to MPTCP" and that systems with no MPTCP listeners may be "vulnerable to similar attacks". As of 1 November 2026, no source confirms or rules out a way to run code without one.

A second path involves request socket migration (net.ipv4.tcp_migrate_req=1 or a BPF_SK_REUSEPORT_SELECT_OR_MIGRATE program). A cloned request then shares the options of the original, and tcp_v4_reqsk_destructor() also freed them with kfree(). The upstream default is 0 (off), and Eric Dumazet wrote that this race "would involve 3 cpus".

IPv6 is not affected by the main issue, because tcp_v6_syn_recv_sock() copies the options. The IPv6 migration case, which needs CALIPSO security labels (an IPv6 option), "will be addressed in a separate patch", says the commit. As of 1 November 2026, that patch is not in the networking tree.

The fix

The commit changes one line in each of two files: inet_sock_destruct() and tcp_v4_reqsk_destructor() now free the options with kfree_rcu(), which waits until every RCU reader has finished. It was merged into mainline on 9 October 2026, first tagged in 7.3-rc7 and backported to every maintained stable line. The kernel CVE team warns: "Cherry-picking individual commits is not recommended or supported by the Linux kernel community at all."

Which systems are exposed

The bug sits in the kernel's handshake code, so the service behind a port does not matter. Red Hat says "Systems accepting inbound TCP connections are most directly exposed." AlmaLinux says that a system "that accepts IPv4 TCP connections from untrusted networks should be treated as exposed". Patch in this order:

  1. Hosts with IPv4 TCP ports open to the internet or other untrusted networks, such as web and SSH servers.
  2. Machines with an MPTCP listener, the setup of Anthropic's exploit. Upstream kernels let programs create MPTCP sockets by default (net.mptcp.enabled = 1), AlmaLinux ships with MPTCP off, and the other vendors' pages do not state their default. Programs built with Go 1.24 or later open MPTCP listeners by default when the main module's go.mod (the Go module file) also names Go 1.24 or later.
  3. Container hosts and Kubernetes nodes, where one kernel serves every container and pod.
  4. Systems with net.ipv4.tcp_migrate_req=1 or a BPF_SK_REUSEPORT_SELECT_OR_MIGRATE program, which open the second path.
  5. Servers reachable only from internal networks. The attack needs only network access to an open IPv4 TCP port (CVSS AV:N), so anyone on those networks can try it.

IPv6 connections do not trigger the main issue, though a host that also accepts IPv4 connections stays exposed through them.

How to check your systems

Run these checks on each host, virtual machine and Kubernetes node. We ran most of them in Docker containers (Docker 29.8.2 on a Mac, with OrbStack 2.2.4); what we saw follows each check.

Find the running kernel

shell
uname -r

This prints the release of the running kernel. As kernel.org notes, "If you see anything at all after the dash, you are running a distribution kernel." Compare such a kernel with your vendor's fixed package in Who is affected. A newer kernel that is installed but not yet booted does not count.

To list the installed kernel packages, and to see whether an update needs a reboot:

shell
# Debian and Ubuntu
dpkg -l 'linux-image*'
# RHEL, AlmaLinux, Rocky Linux, Oracle Linux (RHCK), Amazon Linux
rpm -q kernel
# Amazon Linux 2023 with the 6.12 or 6.18 kernel
rpm -q kernel6.12 kernel6.18
# Oracle Linux with UEK
rpm -q kernel-uek
# dnf systems, after an update: exit code 1 means a reboot is needed
sudo dnf needs-restarting -r

We ran these in Debian 12.15, AlmaLinux 9.8 and Rocky Linux 9.3 containers. Each uname -r printed the kernel of the virtual machine that ran Docker (7.2.9-orbstack-00391-g2b298d05fbd4), as a container has no kernel of its own. The images hold no kernel package: dpkg reported "no packages found matching linux-image*" and rpm printed "package kernel is not installed". needs-restarting -r (from yum-utils) answered "Reboot should not be necessary", which says nothing about the host's kernel. We did not run rpm -q kernel-uek or rpm -q kernel6.12 kernel6.18.

Look for MPTCP listeners

shell
sysctl net.mptcp.enabled
sudo ss -lnMp

net.mptcp.enabled = 1 means programs can create MPTCP sockets; 0 means they cannot. The setting is per network namespace (the separate network stack that a container or pod usually gets). ss -lnMp lists listening MPTCP sockets with the process that owns each one, only in the namespace it runs in. An empty list does not make an unpatched kernel safe: an MPTCP listener is only the setup of Anthropic's exploit.

Our test kernel was built without MPTCP. There, sysctl net.mptcp.enabled failed with "cannot stat /proc/sys/net/mptcp/enabled", and ss printed only its header line with exit code 0, as a kernel with MPTCP and no listener would. The sysctl output tells the two cases apart. We could not test ss against a real MPTCP listener.

Check Go programs

shell
go version -m /path/to/binary
go list -f '{{.DefaultGODEBUG}}' ./path/to/main/package

Since Go 1.24, Go programs open MPTCP listeners by default on Linux. Several things change that: a go.mod (or, in a workspace, a go.work) that names Go 1.23 or older, a godebug multipathtcp=0 line in that file, a //go:debug multipathtcp=0 line in the main package, the GODEBUG=multipathtcp=0 environment variable, and calls to SetMultipathTCP, which override the GODEBUG setting. The build DefaultGODEBUG= line of go version -m lists the defaults compiled into a binary where they differ from the toolchain's: multipathtcp=0 means MPTCP is off by default, and no multipathtcp entry means the toolchain's default applies. go list shows the same for a main package in source form.

We built one small net/http server with go1.27.3 (the golang:latest image) in three modules. go version -m showed multipathtcp=0 for the go 1.23 module and for the go 1.24 module with godebug multipathtcp=0, and no entry for the plain go 1.24 module. Under strace (a tool that logs system calls), only the plain build tried to open an MPTCP socket; our kernel refused it, and the server listened on plain TCP. Started with GODEBUG=multipathtcp=0, that binary made no MPTCP attempt. We did not run go list.

Check the request migration setting

shell
sysctl net.ipv4.tcp_migrate_req

A value of 1 turns on request socket migration, the second path the fix closes.

Check that a firewall rule is loaded

shell
sudo nft list table inet cve_2026_98432   # Red Hat's rule
sudo nft list table ip cve_2026_98432     # AlmaLinux's rule

The counter on the rule line shows how many packets the rule has dropped.

How to fix

Update

Install a kernel that contains the fix, then reboot into it. AWS says "A kernel update and reboot is the only complete remediation for this issue." The kernel CVE team "recommends that you update to the latest stable kernel version for this, and many other bugfixes"; as of 1 November 2026 those are 5.10.273, 5.15.224, 6.1.191, 6.6.161, 6.12.115, 6.18.60 and 7.3.1. Move from 7.2, which reached end of life that day, to 7.3.

On a distribution, install the vendor's fixed package from Who is affected. AWS gives these commands:

shell
# Amazon Linux 2
sudo yum update kernel && sudo reboot
# Amazon Linux 2023
sudo dnf update kernel --releasever=latest && sudo reboot

On Amazon Linux 2023 with the 6.12 or 6.18 kernel, update that package instead, as Amazon's advisories say. Run sudo dnf update kernel6.12 --releasever 2023.12.20261028 or sudo dnf update kernel6.18 --releasever 2023.12.20261026, then reboot.

AlmaLinux asks users to test its fixed kernels "before we release them to production", and suggests turning the testing repository off after the reboot:

shell
sudo dnf install -y almalinux-release-testing
sudo dnf update 'kernel*' --enablerepo=almalinux-testing
sudo reboot
sudo dnf config-manager --disable almalinux-testing

Red Hat, Ubuntu and Rocky Linux have no fixed kernel as of 1 November 2026: install theirs as soon as it ships. SUSE has released some livepatches for SLE 15 SP7 RT. AWS says "applying a livepatch" also resolves the issue but names no package.

If you can't update yet

Red Hat's rule, from RHSB-2026-016, uses nftables (the kernel's packet filtering framework) at the input hook, which sees packets addressed to the machine itself. Save it as cve-2026-98432.nft:

nftables
table inet cve_2026_98432 {
    chain input {
        type filter hook input priority -310; policy accept;
        ip protocol tcp ip hdrlength > 5 counter drop
    }
}

Check it, load it and list it:

shell
sudo nft --check --file cve-2026-98432.nft
sudo nft --file cve-2026-98432.nft
sudo nft list table inet cve_2026_98432

Once the fixed kernel runs, delete it:

shell
sudo nft delete table inet cve_2026_98432

Red Hat says the rule "reduces exposure to this vulnerability but cannot fully prevent it", that filtering "may disrupt necessary workloads and is not a substitute for patching". It asks you to "test it in a non-production environment before deploying it."

We loaded the rule in an AlmaLinux 9.8 container (nftables 1.0.9) and sent single packets from a Debian 12 container with scapy 2.5.0 and hping3 (packet-building tools). Normal TCP connections kept working; each SYN carrying a Record Route or Timestamp option got no reply and raised the rule's counter by one. nft --check in nftables 1.0.6 also accepted the rule. On a container set up as a router, the rule let a Record Route SYN through to the container behind it, and it missed a SYN between two containers on one bridge, in an emulation we built. So on a Docker, Podman or Kubernetes host, the rule does not filter traffic to containers or pods with their own network. We did not test a real Docker or Podman host.

AlmaLinux's command, from its blog post, loads a similar rule at the prerouting hook, which sees every incoming packet before the kernel routes it:

shell
sudo nft -f - <<'EOF'
add table ip cve_2026_98432
delete table ip cve_2026_98432
table ip cve_2026_98432 {
    chain prerouting {
        type filter hook prerouting priority -300; policy accept;
        meta l4proto tcp ip hdrlength > 5 counter drop
    }
}
EOF

AlmaLinux says the rule "blocks the trigger without affecting regular connections", that it "protects the host and the containers that receive their traffic through it, which is the default for Podman and Docker", and gives these limits: "The rule does not survive a reboot." "Traffic between containers on the same host may bypass it." "If the system routes traffic for other machines, forwarded TCP packets with IPv4 options are dropped as well." "Do not use it if your network relies on CIPSO security labels (NetLabel), which are carried in IPv4 options." Remove it with sudo nft delete table ip cve_2026_98432.

Running the command twice in one AlmaLinux 9.8 container left one table with one rule. The probes gave the same results as with Red Hat's rule, and on the router container this rule dropped the forwarded SYN. In the bridge emulation it caught the SYN only when net.bridge.bridge-nf-call-iptables was 1, the setting that sends bridged traffic through these hooks. We did not test firewalld alongside it, a reboot, CIPSO or IPv6.

AlmaLinux also gives steps to turn MPTCP off if you do not need it:

shell
sysctl net.mptcp.enabled
echo 'net.mptcp.enabled = 0' | sudo tee /etc/sysctl.d/90-disable-mptcp.conf
sudo sysctl -w net.mptcp.enabled=0
sudo ss -lnMp

AlmaLinux adds: "This only stops new MPTCP sockets from being created. List the MPTCP listeners that are still open with sudo ss -lnMp and restart the services that own them." AWS gives the same setting and says it "reduces the conditions for running code on a host, but it does not resolve the issue." Our test kernel has no MPTCP, so we could not test these steps.

The AWS bulletin says customers "can disable IP Options to mitigate this with an IP tables rule". This rule drops IPv4 packets of any protocol that carry options, in the INPUT chain only:

shell
iptables -I INPUT 1 -m u32 --u32 '0>>24&0x0f=6:15' -j DROP

For its Network Firewall service, the bulletin gives a stateful rule:

text
drop ip any any -> any any (msg:"CVE-2026-98432: drop IPv4 packets carrying IP options (IHL > 5)"; ipv4.hdr; bsize:>20; classtype:bad-unknown; sid:1000001; rev:2;)

AWS adds that the policy must use strict rule order, with this rule group ahead of your allow rule groups, that traffic your stateless rules pass skips it, and that switching to strict order "can interrupt existing connections". We did not test these two rules.

For Go programs, the Go Security Team gives one setting for the environment, at run time, and one for the main module's go.mod, at build time:

text
GODEBUG=multipathtcp=0
text
godebug multipathtcp=0

Put the first setting in each Go service's environment, then restart the service. With systemd, add Environment="GODEBUG=multipathtcp=0" under [Service] and run sudo systemctl daemon-reload before the restart. After adding the go.mod line, rebuild and redeploy the program.

The Go Security Team warns that "disabling MPTCP on a program-by-program basis is a fragile mitigation. Accidentally leaving a single MPTCP listener exposed will still leave a system vulnerable. Further, it assumes that all viable exploits for this bug depend on MPTCP." Calls to SetMultipathTCP override both settings. In our test, the go.mod line showed up as multipathtcp=0 in go version -m, and under strace a go 1.24 build made no MPTCP call with either setting. We could not test them against a kernel with MPTCP.

Timeline and exploitation status

Times are Dubai time (UTC+4).

WhenWhat
3 Oct 2015 · 15:32David S. Miller commits 079096f103fa ("tcp/dccp: install syn_recv requests into ehash table"), written by Eric Dumazet. The kernel CVE team says the bug starts with it, in Linux 4.4.
25 Sep 2026 · 14:25The Linux kernel CNA reserves CVE-2026-98432.
30 Sep 2026 · 01:43Eric Dumazet posts the first version of the fix on the netdev mailing list, with Xinyang Ge of Anthropic as reporter.
2 Oct 2026 · 02:12The second version of the fix also covers request socket migration, a path the Sashiko AI review bot flagged in the first.
6 Oct 2026 · 03:21Jakub Kicinski commits the fix, e3f33b0a1d89, to the networking tree (netdev/net.git).
11 Oct 2026 · 17:53Linux 6.18.56 and 7.2.10 are announced with the fix.
12 Oct 2026 · 07:22Linux 7.3-rc7 is announced, the first mainline tag with the fix.
15 Oct 2026 · 14:37Linux 6.12.113 is announced with the fix.
15 Oct 2026 · 16:08The Linux kernel CNA publishes CVE-2026-98432, without a CVSS score.
15 Oct 2026 · 16:21Linux 5.10.272, 5.15.223 and 6.1.190 are announced with the fix; 6.6.159 follows at 16:22.
16 Oct 2026 · 17:39Debian publishes DSA-6568-1: trixie is fixed in linux 6.12.113-1.
16 Oct 2026Oracle publishes ELSA-2026-500397, ELSA-2026-500398 and ELSA-2026-500399, which fix UEK6, UEK7 and UEK8.
17 Oct 2026 · 19:39Debian publishes DLA-4841-1: bookworm is fixed in linux 6.1.190-1.
19 Oct 2026 · 03:14Linux 7.3 is released with the fix (18 October 2026 in UTC, the date kernel.org lists).
22 Oct 2026 · 19:09Debian publishes DLA-4850-1: the linux-6.12 package of bookworm is fixed in 6.12.113-1~deb12u1.
24 Oct 2026 · 14:42SUSE releases its first fixed kernels, for SUSE Linux Micro 6.2 (SUSE-SU-2026:24422-1) and SLES 16.0 (SUSE-SU-2026:24446-1). Its CVE page lists them as published on 29 October.
24 Oct 2026 · 18:08The kernel CNA adds its CVSS 3.1 score, 9.8 Critical, to the CVE record.
27 Oct 2026Anthropic's Frontier Red Team publishes its write-up, which describes a root exploit through an MPTCP listener.
30 Oct 2026 · 03:34The Go Security Team posts its advice for Go programs on golang-announce; it reaches oss-security at 04:21.
30 Oct 2026AlmaLinux puts fixed kernels for AlmaLinux 8, 9 and 10 in its testing repository and publishes an nftables rule.
30 Oct 2026Red Hat publishes bulletin RHSB-2026-016, rated Critical, with an nftables rule. The page gives 16:03 as its time, with no time zone.
30 Oct 2026Amazon Linux releases fixes for its Amazon Linux 2 kernels (4.14, 5.4, 5.10 and 5.15) and the Amazon Linux 2023 6.1 kernel. Its 6.18 and 6.12 packages were fixed on 24 and 29 October 2026.
31 Oct 2026 · 09:15AWS publishes security bulletin 2026-160-AWS. Its last update is on 1 November 2026 at 04:15.
1 Nov 2026 · 16:44Linux 7.2.15 is released, and kernel.org marks the 7.2 line end of life.

Exploitation status. As of 1 November 2026, CISA's KEV catalog (version 2026.10.30, checked at 20:53 Dubai time) does not list CVE-2026-98432, and none of the CVE, vendor and research pages under Sources reports attacks. Anthropic's Frontier Red Team built a working exploit and described it on 27 October 2026, in a write-up that links no code for it. Its disclosure policy says that once a patch is out, it "would generally wait 45 days before publishing full technical details"; it has given no date for this flaw. A GitHub search at 20:40 Dubai time on 1 November 2026 returned 40 public repositories that mention the CVE, all created from 29 October 2026 on. Going by their descriptions, seven describe a proof of concept or a trigger for the race, one of them claims to recover KASLR offsets, and none claims remote code execution. We did not open or run their code, and we did not try to trigger the flaw: each of our test probes was one packet with no retransmission.

Updates to this article

We revise this article when Red Hat, Ubuntu or Rocky Linux ships a fixed kernel or AlmaLinux moves its kernels to production, when CISA lists the CVE as exploited or a vendor reports attacks, when NVD publishes its analysis, when the IPv6 follow-up patch lands, when Go decides on proposal 82717 to turn MPTCP off by default, or when exploit code or a way to run code without an MPTCP listener is published.

  • 1 November 2026: Rewritten in full. One correction: Red Hat's bulletin says nothing about containers; the container limit of its rule comes from our own tests (AlmaLinux says traffic between containers may bypass its own rule). Added: fixes from Amazon (4.14 and 5.4 kernels) and Oracle (5.4-based UEK6) for lines older than 5.10, where upstream fixes stop; the AWS bulletin, the end of the 7.2 line, the fixes from Oracle, SUSE, Amazon Linux and AlmaLinux (testing), other vendors' ratings, public repositories and our container tests.

Get help

Patching starts with the list of Linux systems that accept IPv4 TCP connections, cloud instances and container hosts included. An external test of your internet-facing addresses shows which hosts answer on which TCP ports and which software versions they reveal. A review of server and container settings shows where MPTCP is on, which Go services open MPTCP listeners and whether request migration is enabled, which helps you decide what to patch first and where a vendor's firewall rule fits.

External security audit

Your domains and public IP addresses, checked from the internet: open ports, exposed files, email spoofing, encryption and known flaws, graded from A to F.

From AED 15,000, excl. 5% VAT

Ask for a quote Book a call

Sources

More CVE alerts

All CVE alerts