Information Security · Linux Kernel · Operating Systems · AI Capability · News AnalysisCVE of 15 October 2026, record to 1 November 2026
Patched upstream, waiting downstream
A race in Linux's TCP handshake, open since kernel 4.4
On 15 October 2026 the Linux kernel's CVE team published CVE-2026-98432, a race in the code that sets up incoming IPv4 TCP connections. Since Linux 4.4 the kernel has handed a connection's IP options to the new socket without copying them, so a retransmitted SYN handled on another CPU can read them after they are freed. The kernel team scores it 9.8 out of 10: reachable over the network, with no login and no user action. The fix is in every supported stable kernel. What is uneven is everything after that: some distributions shipped fixes within a day, while Red Hat and Ubuntu had none by 1 November, and the most alarming claim about the bug, a remote root exploit, has been published but not reproduced.
What Happened
One freed pointer, reachable from the network Graded by how far each item is established
- 9.8CVSS 3.1 score from the kernel CVE team, matched by Red Hat
- 4.4The kernel release that introduced the bug
- 8Kernel lines with a fix, from 5.10.272 to 7.3
- 0Confirmed exploitation in the wild, per CISA's catalogue of 30 October
Confirmed
The mechanism, the 9.8 score and the fixed versions come from the kernel CVE team's own announcement. Red Hat's CVE page also shows 9.8. CISA's Known Exploited Vulnerabilities catalogue, version 2026.10.30, does not list the flaw.
Claimed, and disputed
Anthropic says an unreleased internal model wrote a working remote root exploit against servers with a multipath TCP listener; it has not published the code. CloudLinux says its kernel team could not reproduce the claim.
Timeline
From kernel 4.4 to a crowded week Dates as each source gives them
- Linux 4.4A change that lets request sockets be processed without the listener lock opens the race (kernel CVE team).
- 15 Oct 2026CVE published with fixes named in eight kernel lines (kernel CVE team; NVD).
- 16 Oct 2026Debian trixie and Oracle Linux's UEK kernels ship fixes (The Hacker News).
- 24 Oct 2026Amazon Linux 2023's kernel6.18 is fixed (Amazon).
- 27 Oct 2026Anthropic publishes its account of a model-written remote root exploit against multipath TCP listeners (Anthropic).
- 29 Oct 2026Go's security team warns that Go 1.24 and later open listening sockets as multipath TCP by default (oss-security).
- 30 Oct 2026Red Hat issues a Critical bulletin with fixes still pending; AlmaLinux puts patched kernels in testing; CloudLinux says it could not reproduce the exploit; BruCERT issues an alert; Amazon fixes its remaining kernels (Red Hat; AlmaLinux; CloudLinux; BruCERT; Amazon).
The Argument
How bad depends on the race; how exposed depends on the distribution What the record supports, and what complicates it
The vendors agree on what the bug is and disagree on how hard it is to use. The kernel CVE team and Red Hat score it 9.8, treating the race as easy to win. Amazon scores it 8.1 with high attack complexity, and The Hacker News reported on 29 October that Red Hat, Amazon, Oracle and SUSE then sat between 5.9 and 8.1. The difference is one letter in the vector, and Anthropic's account bears on it directly: losing the race only makes the handshake fail, so an attacker can try again, and its proof of concept took 48 minutes to leak a kernel pointer and five more to run commands. Read that way, a race that can be retried indefinitely is not much of a barrier. That reading is ours, and it rests on an exploit nobody outside Anthropic has seen. CloudLinux says it could not reproduce it, and no one has shown remote code execution on a server without a multipath TCP listener. The Go team's warning cuts the other way: the bug lives in ordinary IPv4 TCP code, not in multipath TCP.
- 48 + 5 minAnthropic's stated time to leak a kernel pointer, then run commands as root
- 9.8 / 8.1The kernel team's and Red Hat's score against Amazon's
What complicates it
The exploit account comes from the company whose model wrote it, with no code, kernel version or configuration published. A failed reproduction by one vendor is not proof the exploit does not work, and a successful one in a lab would not show it works outside one. Until someone independent publishes either, the 9.8 rests on the kernel team's reading of the race rather than on a demonstrated attack.
What Others Add
Who is fixed, and what to do meanwhile Status as each vendor's own page showed it on 1 November 2026
| Distribution or question | What the record shows | Source |
|---|---|---|
| Debian, Oracle Linux | Fixed on 16 October: Debian trixie in linux 6.12.113-1 (DSA-6568-1), Oracle's UEK kernels in ELSA-2026-500397 to 500399. Debian 12 followed through its LTS channel. | The Hacker News |
| Amazon Linux | Every listed line fixed: kernel6.18 on 24 October, kernel6.12 on 29 October, and the default Amazon Linux 2023 kernel and Amazon Linux 2 on 30 October. | Amazon |
| SUSE | Fixed for some products, including SLES 16.0 and SUSE Linux Micro 6.2, from kernel-default 6.12.0-160000.40.1. | The Hacker News |
| Red Hat, AlmaLinux | RHEL has no fixed kernel; Red Hat's bulletin is Critical and ongoing and says fixes are being expedited. AlmaLinux 8, 9 and 10 have patched kernels in their testing repository. | Red Hat; AlmaLinux |
| Ubuntu | No fix released. Work is in progress from 22.04 onward; 20.04, 18.04 and 16.04 are listed as vulnerable. Priority High. | Ubuntu |
| Why do Go services matter? | Since Go 1.24, Go programs on Linux open listening sockets as multipath TCP by default, the condition Anthropic's exploit needs. GODEBUG=multipathtcp=0 turns it off, though the Go team calls per-program fixes fragile. KrakenD says its gateway matches the conditions on unpatched hosts. | oss-security; KrakenD |
| What stops it before a fix? | An nftables rule dropping incoming IPv4 TCP packets that carry IP options. Ordinary TCP options sit in the TCP header and are untouched. Red Hat says the rule reduces exposure but cannot fully prevent exploitation; BruCERT warns that disabling multipath TCP alone is not a complete mitigation. | Red Hat; AlmaLinux; BruCERT |
| Can a host patch without rebooting? | SiteGround says it built and applied its own live patch across its servers the same day. CloudLinux says a KernelCare live patch is in preparation. | SiteGround; CloudLinux |
Conclusion
Patch where you can, filter where you cannot What the record settles, and what to watch
Settled
The bug is real, remotely reachable and fixed in every supported upstream kernel. It is not listed as exploited in the wild. Debian, Oracle, Amazon and parts of SUSE have shipped fixes.
Open
When Red Hat and Ubuntu ship fixed kernels; whether anyone independent reproduces the remote root exploit; and whether a way in exists on servers with no multipath TCP listener.
- Patch and reboot where a fix exists. A kernel update does nothing until the new kernel is running, unless a live patch names this CVE.
- Elsewhere, drop IPv4 TCP with IP options. Treat it as a stopgap, not a fix, and do not use it where CIPSO labelling depends on IP options.
- A container image cannot fix it. Containers share the host's kernel; the node has to be patched.
Quick notes
- Greg Kroah-Hartman for the Linux kernel CVE team, “CVE-2026-98432: ipv4: free inet_opt and ireq_opt after an RCU grace period”, 15 Oct 2026
- NIST, National Vulnerability Database record, 15 Oct 2026
- Red Hat, CVE page, and RHSB-2026-016, 30 Oct 2026
- Canonical, Ubuntu CVE page
- Andrew Lukoshko, AlmaLinux OS Blog, 30 Oct 2026
- Amazon Web Services, Amazon Linux Security Center
- Anthropic Frontier Red Team, “Discovering and exploiting vulnerabilities in critical, real-world software with Claude”, 27 Oct 2026
- Go security team on oss-security, 29 Oct 2026
- CloudLinux Blog, 30 Oct 2026
- CISA, Known Exploited Vulnerabilities Catalog, version 2026.10.30
- Swati Khandelwal, The Hacker News, 29 Oct 2026
- Lawrence Abrams, BleepingComputer, 30 Oct 2026
- KrakenD Blog, 30 Oct 2026
- BruCERT, 30 Oct 2026
- SiteGround Blog, 30 Oct 2026. All read 1 Nov 2026.