If you run Ubuntu 22.04, 24.04, or 26.04 on anything that accepts inbound TCP connections, the practical answer is short: check your running kernel against Ubuntu’s CVE tracker rather than the version string, count the TCP ports the internet can reach, apply the interim controls that Red Hat and AlmaLinux have actually published, and be ready to reboot the day Ubuntu’s fixed kernels leave testing. As of November 5, those fixes were still sitting in Ubuntu’s -proposed pocket with the tracker reading “Vulnerable” for every LTS release, roughly three weeks after upstream shipped the patch, according to AtlasSecurity’s detailed write-up. Community reports suggest packages were already appearing in apt while the CVE page lagged behind, so the first step below may already resolve itself. Check before you do anything else.
The reason this deserves your weekend is the shape of the bug, not just the score. CVE-2026-98432 is a use-after-free in the IPv4 TCP handshake, reachable by a remote, unauthenticated peer, rated CVSS 3.1 9.8 by the kernel’s own CVE Numbering Authority and, independently, by Red Hat. The trigger is unusual: it requires the connecting peer’s IPv4 header to carry options, and the one detailed public description, Anthropic’s, requires a multipath TCP (MPTCP) listener. Those preconditions cut both ways. They mean a well-filtered host is harder to hit than the 9.8 implies, and they mean you can cheaply remove one of the preconditions while you wait.
How the IPv4 handshake use-after-free works
The mechanism, as AtlasSecurity explains it, is a lifetime bug in half-open connection handling:
When a Linux server accepts a TCP connection, the kernel tracks the half-open handshake in a lightweight “request socket,” then creates a full socket once the handshake completes. If the connecting peer’s IPv4 header carried options, the kernel keeps them to echo in its replies. The bug: the kernel handed those saved options from the request socket to the new connection without copying them, then freed them immediately when the connection went away, while another CPU could still be reading them to build a handshake reply.
Three properties follow. The buggy code runs before any application sees the connection, so application-layer hardening does nothing. No credentials are needed. And the bug is old, present since Linux 4.4, which covers every supported Ubuntu LTS kernel. Upstream fixed it with a four-line change in stable releases 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, and 7.2.10 on October 11 and 15, and in 7.3 on October 18.
On exploitability, the strongest claims come with labels attached. Anthropic, whose researcher reported the bug, says its Claude model built a working proof of concept that runs commands as root on “Linux servers with a multipath TCP listener,” with no credentials; that is the reporting vendor describing its own lab result, not an independent reproduction. Separately, the write-up relays a security research team’s November 3 report of scanning on port 443 over MPTCP, strongly suspected to target this CVE, and that team’s count of about 2.1 million servers reachable over MPTCP on port 443. Both figures are single-source. Treat them as reasons to hurry, not as proof your host is being attacked tonight.
First check: are you actually exposed?
The verification guidance in the write-up is explicit about three traps that produce wrong answers:
Compare the result with your distribution’s status above, not with the upstream version. If the installed kernel is newer than the running one, the fix isn’t active until you reboot. Run these on container hosts and Kubernetes nodes, not inside containers:
Each of those deserves a sentence. Ubuntu kernels carry backported fixes under Ubuntu’s own version numbers, so comparing uname -r output against upstream numbers like 5.15.223 will mislead you in both directions; the source of truth is Ubuntu’s CVE tracker and the eventual Ubuntu Security Notice. An installed-but-not-running kernel protects nothing, which matters on fleets where configuration management installs updates but reboots wait for a maintenance window. And a container reports the host’s kernel, so a version check inside one does describe the exposure; the comparison that breaks is against a distribution’s status page, because the container’s userland may be a different distribution entirely. The checklist belongs on the host or node.
The second half of exposure is reachability. The bug sits in code that runs for any listening TCP port, so the inventory question is simply: which TCP ports on this host can an untrusted peer connect to? AWS’s bulletin, quoted in the write-up, states that hosts behind Application, Classic, or Network Load Balancer TLS listeners, CloudFront, or API Gateway “do not receive TCP connections from untrusted sources directly when following best practices,” while still recommending patching those hosts within 30 days. Its November 3 update adds that Global Accelerator drops packets carrying IP options entirely. If your hosts sit behind such a layer, your exposure is real but attenuated; if they terminate TCP directly, they are the target profile.
One counterweight, offered by an Ask HN commenter with firewall visibility:
I haven’t had a documented packet with IPOptions arrive on any of my firewalls, the packets are typically dropped, so exploiting remotely is going to be very difficult.
That is one anonymous anecdote, and the same thread’s corporate edge may drop options that yours passes. It is a reason to prioritize by exposure, not a reason to skip patching.
The MPTCP question, and why Go services change your answer
The reported exploit path needs an MPTCP listener. Here the news is uncomfortable for a specific population: per the Go Security Team’s notice, Go programs have opened MPTCP listeners by default since Go 1.24, a default that follows the go line in the program’s go.mod and applies only where the kernel allows MPTCP (net.mptcp.enabled = 1; AlmaLinux says it ships MPTCP off). A Go service that never asked for multipath TCP may still present the listener the exploit wants. So the exposure inventory has a second question alongside open ports: does this host run Go services, and do they listen on reachable addresses?
Two cautions come from Red Hat’s bulletin RHSB-2026-016, as relayed by AtlasSecurity: checking for MPTCP listeners “is not a test for exposure,” and turning MPTCP off “only blocks one method of exploitation.” In other words, absence of a listener does not clear a host, and removing the listener removes the documented exploit’s setup, not the bug. The underlying use-after-free remains reachable through ordinary TCP with IPv4 options.
Interim controls the sources actually document
This is where discipline matters, because the controls that exist are narrower than the anxiety. The documented options, per the write-up’s October 31 and later updates, are:
- The packet-filter rule from Red Hat bulletin RHSB-2026-016 per the write-up (published October 30, when Red Hat raised its rating from Important 8.1 to Critical 9.8), with an iptables variant. The iptables form needs the kernel’s
xt_u32module, which the Debian 12, Ubuntu 22.04 and 24.04, and RHEL 8 kernels include; RHEL 9 and 10 kernels leave it out, so the nftables form applies there. The write-up’s version hooks prerouting rather than input, which also covers packets the host routes onward to containers. The rule lasts until reboot, though Ubuntu’s own tracker publishes a rule that does the same thing, with steps to load its version at boot from /etc/nftables.conf, per the write-up. I am deliberately not printing the syntax here: get it from the bulletin or the write-up, not from a paraphrase. - Turn off MPTCP if you do not use it, following AlmaLinux’s published steps. New MPTCP sockets then fail and Go programs fall back to plain TCP at start. Again, this removes the reported exploit’s precondition, not the vulnerability.
The limits are stated, and they bite. AlmaLinux warns that “Traffic between containers on the same host may bypass it,” and the write-up’s own namespace testing confirmed a Record Route SYN bridged between namespaces never hit the rule. If your threat model includes east-west traffic between containers on one host, the filter is not your control. AlmaLinux also says not to use the rule on networks that rely on CIPSO security labels, which legitimately travel in IPv4 options; dropping all option-carrying packets there breaks labeled networking.
Beyond these, nothing in the sourced material documents a safe mitigation. If you see a sysctl or netfilter recipe circulating without a bulletin behind it, treat it as unverified.
How the timeline actually played out
Reconciling the per-distro dates from AtlasSecurity’s running status log and the October 31 news coverage:
| Date (2026) | Event |
|---|---|
| Oct 11 and 15 | Upstream stable kernels ship the four-line fix |
| Oct 16 | Oracle Linux fixed in UEK kernels |
| Oct 27 | Anthropic’s Frontier Red Team describes the Claude-built remote root exploit |
| Oct 28 | Ubuntu 26.10’s fixed kernel 7.3.0-10.10 reaches its updates pocket |
| Oct 29 | Go Security Team: Go opens MPTCP listeners by default since Go 1.24 |
| Oct 30 | Red Hat raises rating to Critical 9.8, publishes RHSB-2026-016 mitigation; Amazon Linux completes its fixes |
| Nov 1 | Red Hat ships fixed kernels for RHEL 8, 9, and 10 |
| Nov 2 | AlmaLinux ships fixed kernels |
| Nov 3 | Reported MPTCP scanning of port 443; AWS updates its bulletin |
| Nov 5 | Ubuntu 22.04/24.04/26.04 fixes still in -proposed; tracker reads “Vulnerable”; SUSE only partly fixed (SLES 15 SP7 done) |
Two fairness notes. First, the popular claim that Ubuntu was uniquely laggard is not quite right: the same status board listed Red Hat and SUSE as only partly fixed on November 5 (Red Hat’s RHEL CoreOS for OpenShift was still affected, for instance). Second, the gap may already be closed by the time you read this; an HN commenter reported the fixed packages were “showing in apt” while Ubuntu’s CVE page stayed stale. Timestamp everything you check.
Wait, mitigate, or move: a decision framework
The honest framing of the choice comes from the October 31 coverage: “affected administrators are left to decide whether to stay on the stock kernel, apply out-of-tree mitigations if any exist, or move workloads onto releases or kernels that already carry the patch.” Here is how I would sequence it, given the evidence:
- Verify first, always. Running kernel versus tracker, reachable TCP ports, MPTCP listeners. A few minutes per host class.
- If you can filter inbound TCP and reboot soon: apply the RHSB-2026-016 rule (mind the container and CIPSO limits, documented in the write-up), disable MPTCP if unused, and plan to install and reboot into Ubuntu’s kernel the day it leaves
-proposed. With fixes already in testing, this is a short-horizon wait, and the documented controls shrink the target. - If you cannot filter or cannot reboot promptly: the interim controls will not hold the line for long against a bug this reachable, and the write-up’s scanning report raises the cost of waiting. Moving exposed workloads to an already-patched kernel (Ubuntu 26.10, a fixed RHEL/Alma/Oracle/Amazon kernel, or a load-balancer-fronted posture that blocks direct untrusted TCP) is the defensible choice, at real migration cost.
- Backporting the four-line upstream patch yourself is technically small and organizationally large: you take over kernel QA and maintenance for a fork. I would only consider it for a host that cannot be filtered, rebooted, or moved, and even then the in-testing Ubuntu package is likely to land first.
On prioritization generally: CVSS 9.8 is one input, not the verdict. Research on vulnerability scoring documents that CVSS “may not provide a precise view of the risks related to software vulnerabilities,” even as compliance regimes lean on it, and research on exploit-likelihood scoring argues for ranking by estimated exploitation likelihood. Here both signals point the same way, high score plus a public exploit description plus reported scanning, so the conditional part of your judgment is exposure, not urgency.
Why a four-line fix took three weeks
The structural answer is in the kernel’s maintenance economics. Research on patch porting across the kernel ecosystem finds that stable and LTS branches integrate only bug fixes from mainline under predefined rules, that some LTS branches are maintained for six years, and that distribution kernels are customized forks of those branches. About 40% of LTS commits arrive without stable tags, per the same research, and get routed through the AUTOSEL neural-network bot plus manual maintainer review. A four-line mainline fix therefore traverses several queues before it reaches your apt mirror: stable selection, distro backport, distro build, distro QA. An HN commenter pointed to Ubuntu’s kernel QA process and a September post about shortening update timeframes, concluding the pipeline “aren’t flexible enough to get a 4-line fix out the door when they need to.”
The second-order point is uncomfortable for fleets standardized on Ubuntu LTS precisely because it is boring. If AI-assisted discovery keeps compressing the time from bug report to working exploit, as the Claude-built PoC here at least suggests, the window in which “wait for the distro” is a complete strategy shrinks, and the interim-mitigation burden shifts to you. That is an inference from one incident and vendor-reported capability, not a measured trend. But the verify-and-mitigate muscle you build this week is the asset that survives it.
What to re-verify before you act
Every patch-status claim in this piece is secondhand and timestamped November 5-6; treat each status as provisional until you have confirmed it against the primary source, especially since the tracker page itself stayed stale while apt moved, per one HN commenter. Before acting, check Ubuntu’s CVE tracker and any USN for your release, pull the RHSB-2026-016 bulletin for the actual rule syntax (linked from the write-up), and confirm whether the 22.04/24.04/26.04 kernels have left -proposed. If they have, the decision framework above collapses to its last step: install, reboot, verify the running kernel, and note which of your hosts could not have rebooted quickly, because that list is your real exposure register for the next one.
Frequently Asked Questions
What interim controls are documented for mitigating CVE-2026-98432?
- The packet-filter rule from Red Hat bulletin RHSB-2026-016 per the write-up (published October 30, when Red Hat raised its rating from Important 8.1 to Critical 9.8), with an iptables variant. The iptables form needs the kernel’s
xt_u32module, which the Debian 12, Ubuntu 22.04 and 24.04, and RHEL 8 kernels include; RHEL 9 and 10 kernels leave it out, so the nftables form applies there. The write-up’s version hooks prerouting rather than input, which also covers packets the host routes onward to containers. The rule lasts until reboot, though Ubuntu’s own tracker publishes a rule that does the same thing, with steps to load its version at boot from /etc/nftables.conf, [per the write-up](https://atlasum.net/blog/cve-2026
Does running Go services increase the risk of CVE-2026-98432?
per the Go Security Team’s notice, Go programs have opened MPTCP listeners by default since Go 1.24, a default that follows the go line in the program’s go.mod and applies only where the kernel allows MPTCP (net.mptcp.enabled = 1; AlmaLinux says it ships MPTCP off). A Go service that never asked for multipath TCP may still present the listener the exploit wants.

Join the discussion
Share a useful perspective or ask a question about this article.