Go straight to an AI summary of this blog post instead

  • Summarize this post with Perplexity
  • Summarize this post with Gemini

KernelCare Rebootless Patch for CVE-2026-98432 Is Out for Debian 13

KernelCare’s rebootless patch for CVE-2026-98432 is out for Debian 13. Work on Debian 12 and Oracle Linux UEK is in progress. Rebootless patches for EL 8, 9 and 10 (RHEL, AlmaLinux, CloudLinux, Oracle Linux RHCK and Rocky Linux) are being built now, without waiting for Red Hat, which still hasn’t released a fixed kernel for RHEL 8, 9 or 10. Ubuntu patches aren’t available yet and have no release date.

KernelCare status (October 29, 2026): RHEL 7 and CentOS 7 kernels aren’t affected. LibCare isn’t involved, since this is a kernel bug. Until a rebootless patch ships for your kernel, the mitigations further down can reduce exposure on affected hosts.

Get KernelCare →

Key Takeaways

  • CVE-2026-98432 lets a remote attacker trigger a use-after-free during the Linux IPv4 TCP handshake.
  • Anthropic says an unreleased Claude model turned it into a remote root exploit against MPTCP listeners.
  • Upstream fixed it on October 5, but RHEL-based kernels and Ubuntu still lack fixes.
  • Go 1.24 and later enable MPTCP on listening sockets by default, but EL 8, 9 and 10 keep MPTCP switched off.

The Short Version

Jump to the Long Version ↓

What happened: On October 15, 2026, the kernel CVE team published CVE-2026-98432, a use-after-free race in the code that handles IPv4 TCP handshakes, and added a CVSS 9.8 score on October 24. The two-line fix was applied upstream on October 5 and is in every maintained stable branch. Anthropic researcher Xinyang Ge reported it. Jump to this section →

The claim: On October 27, Anthropic said an unreleased, internal Claude research model wrote a working exploit that lets a remote attacker with no credentials run commands as root on Linux servers with a Multipath TCP (MPTCP) listener. The exploit code, the kernel version and the tested configuration aren’t public, and no independent public reproduction had turned up as of October 30. Jump to this section →

Who’s affected: Any kernel from Linux 4.4 onward that hasn’t picked up the fix, on a server that accepts IPv4 TCP connections. That includes RHEL 8, 9 and 10 and the distributions built on them, plus Ubuntu. RHEL 7 and CentOS 7 aren’t affected, and IPv6 connections don’t trigger the main bug. Jump to this section →

Why MPTCP matters: Go programs built with Go 1.24 or later enable MPTCP on their listening sockets by default. EL 8, 9 and 10 ship with MPTCP switched off, though, so a stock EL host can’t open an MPTCP listener unless someone turned it on. The Go Security Team still says systems without an MPTCP listener may be vulnerable to similar attacks. Jump to this section →

Where things stand: This isn’t a zero day. The fix came before the CVE, Anthropic hasn’t published its exploit code, and as of October 30 no exploitation in the wild has been reported and the bug isn’t in CISA’s KEV catalog. Debian, Oracle UEK, most supported SUSE releases and some Amazon Linux kernels are fixed. Red Hat and Ubuntu haven’t shipped a fixed kernel and haven’t given a date. KernelCare’s rebootless patch is out for Debian 13, and patches for EL 8, 9 and 10 are being built without waiting for Red Hat. Jump to this section →

What to do today: Find out whether anything on your servers listens on MPTCP. With -H dropping the header row, every line this command prints is an MPTCP listener:

ss -H -lMn

On hosts with no fixed kernel, dropping inbound TCP packets that carry IP options should block the documented trigger, MPTCP or not. It’s untested, so try it in staging first. Jump to this section →

The Long Version

Most network-reachable kernel bugs come with a catch. A module has to be loaded, a niche protocol switched on, or a service running that hardly anyone uses. CVE-2026-98432 sits in the code that accepts IPv4 TCP connections. If a server takes IPv4 TCP connections at all, it runs this code.

That alone would make a 9.8 worth a look. What pushed it into the headlines is Anthropic’s claim that one of its models turned it into a remote root exploit with no credentials needed. The CVE itself had gone out quietly twelve days before that, one of the many the kernel CVE team publishes every week.

The awkward bit is the timeline. Upstream fixed this on October 5. Debian had a fixed kernel out by October 16. Red Hat still has none for RHEL 8, 9 or 10, and that gap, more than the AI angle, decides how long affected enterprise systems stay exposed.

How a TCP Handshake Ends Up Reading Freed Memory

The trigger is a TCP SYN carrying IPv4 options, specifically Record Route or Timestamp. Normal traffic almost never sends these, but nothing stops a client from doing it. The listening kernel saves those options with the pending connection request, in ireq_opt.

When the handshake completes, tcp_v4_syn_recv_sock() hands the options to the new child socket without making a copy, so the request socket and the child now point at the same memory. If the client resends its SYN at the right moment, another CPU can pick up that retransmitted SYN and start building a SYNACK from the same options in ip_options_build(). If the child socket is torn down in the meantime, inet_sock_destruct() frees the options with a plain kfree(), and the SYNACK path reads memory that’s already gone.

The bug dates back to commit 079096f103fa (“tcp/dccp: install syn_recv requests into ehash table”), which let request sockets be processed without holding the listener lock. That commit shipped in Linux 4.4 in January 2016. Red Hat lists RHEL 7 as not affected, and CentOS 7 runs the same kernel line.

The fix from Eric Dumazet is about as small as kernel fixes get. One kfree() in inet_sock_destruct() in net/ipv4/af_inet.c and one in tcp_v4_reqsk_destructor() in net/ipv4/tcp_ipv4.c become kfree_rcu(), so anything still reading the options finishes before the memory is freed. The first one is the free that bites in the main path, and the second covers the migration path below.

There’s a second, narrower path through request socket migration, which applies only with net.ipv4.tcp_migrate_req=1 or a BPF_SK_REUSEPORT_SELECT_OR_MIGRATE program attached. Both are off by default. If you’ve read that this bug “needs a non-default configuration,” that’s only true for the migration path. The main path needs none.

IPv6 dodges the main path because tcp_v6_syn_recv_sock() copies the options. The CVE record also flags a similar migration issue on the IPv6 side, limited to options set by CALIPSO, and upstream plans a separate patch for that one.

The race is hard to win, but that’s less reassuring than it sounds. A lost race just fails that one handshake. It doesn’t crash the server, so an attacker can keep trying for as long as the port stays open.

What Anthropic Says Its Model Built

Anthropic described the exploit in an October 27 post from its Frontier Red Team, alongside exploit chains for Chrome, Safari, Android and Windows. The post says the work used an unreleased, internal research model, and that the model wrote the Linux exploit after “initial prompting.”

Anthropic describes two stages. First, winning the race makes the kernel read an attacker-controlled length field and copy extra bytes into the SYNACK it sends back, and once one of those replies carries a kernel pointer, KASLR is defeated. Second, the same bug overflows a buffer and overwrites a callback pointer with the address of a payload already staged in kernel memory. When the kernel calls that overwritten pointer, a shell command of the attacker’s choice runs as root. Anthropic says the address leak took 48 minutes in its test, and command execution took another five.

TuxCare hasn’t reproduced the exploit, and no independent public reproduction had turned up as of October 30. Anthropic didn’t publish the code, the kernel version or the configuration it tested. The post also doesn’t say whether servers without an MPTCP listener can be taken over the same way.

The claim isn’t out of line with what the kernel CVE team already wrote, though. The team’s own scoring notes in the CVE record describe the stale options leaking adjacent kernel heap memory back to the peer and steering out-of-bounds kernel writes. Anthropic’s chain is basically that description put to work.

Why MPTCP Listeners Are More Common Than You’d Think

MPTCP lets one connection use several network paths at once. On Linux, an application only gets it by asking for it when it opens a socket, so for years most teams could safely assume they weren’t running it.

Go changed that. Since Go 1.24, Go programs on Linux enable MPTCP on listening sockets by default where the kernel supports it. A Go-based API gateway, reverse proxy or internal service built with Go 1.24 or later may be holding an MPTCP listener its owners never asked for.

EL shops get some good news here. Red Hat ships RHEL with MPTCP support disabled, and the rebuilds ship the same way, so turning it on means setting net.mptcp.enabled=1 yourself. On a stock EL 8, 9 or 10 host, a Go service asks for MPTCP, the kernel says no, and the service quietly falls back to a plain TCP listener. Unless someone on your team switched MPTCP on, the chain Anthropic demonstrated has nothing to aim at.

On October 29, the Go Security Team said setting GODEBUG=multipathtcp=0 in a program’s environment turns that default off at runtime, and adding godebug multipathtcp=0 to the main module’s go.mod does the same at build time. Code that calls SetMultipathTCP on a net.ListenConfig or net.Dialer overrides both. The team also called this a fragile mitigation, since a single MPTCP listener left exposed still leaves a system vulnerable. And it went a step further. The underlying bug isn’t specific to MPTCP, and systems with no MPTCP listeners may be vulnerable to similar attacks.

Plan around that last point anyway. Nobody has publicly shown root against a server without an MPTCP listener, and the demonstrated chain needs one. But the bug itself is reachable through any IPv4 TCP listener, and the Go team won’t rule out other routes, so a host with MPTCP off is less exposed, not fixed.

If sysctl net.mptcp.enabled returns 1, applications on that host can open MPTCP sockets. On EL 8, 9 and 10, 0 is the default. On a reasonably recent iproute2, ss -H -lMn lists the MPTCP listeners that actually exist, without the header row. If ss rejects the -M flag, that host’s iproute2 is too old for this check.

Why Vendors Can’t Agree on a Score

The kernel CVE team scores CVE-2026-98432 at 9.8, with network attack vector, no privileges and low attack complexity. Its reasoning is that both sides of the race are “attacker-induced,” since the attacker controls the SYN, the retransmission and the connection teardown. Ubuntu’s page shows the same 9.8. NVD hadn’t added its own score as of October 30.

Red Hat started out much lower. Its CVE page first rated the flaw Moderate (5.9), with a crash as the most likely outcome and no demonstrated confidentiality or integrity impact. On October 29, its machine-readable advisory file moved to Important (8.1) with high confidentiality and integrity impact, and by October 30 the CVE page showed the same 8.1. Oracle, SUSE and Amazon also score it 8.1.

All of that remaining spread comes down to one metric. The 8.1 vector shared by Red Hat, Oracle, SUSE and Amazon matches the kernel team’s on everything except attack complexity, which those vendors rate high. That’s a fair technical point about a race. It matters less when losing the race costs the attacker nothing but one connection attempt.

The practical problem is tooling. If your scanner or patch SLA keys off the vendor’s severity, this flaw may have shown up on EL hosts as Moderate, or not as urgent at all, for most of October, despite a publicly described remote root chain. If your process tiers patches by vendor severity, override it for this CVE.

Which Distributions Have a Fix

Distribution Status as of October 30, 2026 Fixed kernel or advisory
Upstream stable Fixed 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3
RHEL 8, 9 and 10 Affected, no date None released
AlmaLinux, CloudLinux, Rocky Linux and Oracle Linux RHCK 8, 9 and 10 Affected None released
RHEL 7 and CentOS 7 Not affected Not needed
Oracle Linux UEK Fixed October 16 ELSA-2026-500397, ELSA-2026-500398, ELSA-2026-500399
Debian 13 (trixie) Fixed linux 6.12.113-1 (DSA-6568-1)
Debian 12 (bookworm) Fixed linux 6.1.190-1 (DLA-4841-1), linux-6.12 6.12.113-1~deb12u1 (DLA-4850-1)
Ubuntu 22.04, 24.04, 26.04 and 26.10 Fix in progress None released
Ubuntu 16.04, 18.04 and 20.04 Vulnerable None released
Amazon Linux 2023 kernel6.18 and kernel6.12 packages Fixed ALAS2023-2026-3251, ALAS2023-2026-3335
Amazon Linux 2 and the default Amazon Linux 2023 kernel package Pending None released
SLES 16.0, SLE Micro 6.2 and openSUSE Leap 16.0 Fixed See SUSE’s CVE page
SLES 15 SP4 to SP6 LTSS and SLE Micro 5.3 to 5.5 Fixed See SUSE’s CVE page
SLES 15 SP7 Affected, fix in progress None released

CloudLinux 9 and 10 run AlmaLinux kernels, so they follow the AlmaLinux row. Red Hat lists the issue for RHEL 8, 9 and 10 but gives no release date, and Ubuntu doesn’t give one either.

If you build your own kernels, the fix is in the stable releases above. The kernel CVE team recommends moving to the latest stable kernel rather than cherry-picking individual fixes.

Buying Time Until a Rebootless Patch Lands

As of October 30, Red Hat says no available mitigation meets its standards, and no other vendor advisory lists one. The options below follow from how the bug works, but none of them has been tested against an exploit in-house. Try each on a staging host first, and don’t treat any of them as a substitute for a fixed kernel.

The option with the widest reach is dropping inbound TCP packets that carry IPv4 options. The main path needs IP options in the SYN, normal TCP traffic doesn’t carry them, and this rule doesn’t care whether anything on the host uses MPTCP:

nft add table ip cve_2026_98432
nft add chain ip cve_2026_98432 prerouting '{ type filter hook prerouting priority -300; }'
nft add rule ip cve_2026_98432 prerouting meta l4proto tcp ip hdrlength != 5 drop

To undo it, run nft delete table ip cve_2026_98432. The rule doesn’t survive a reboot, so add it to your persistent firewall configuration if you decide to keep it. Filtering IP options at the network edge can cover every host behind it, as long as all traffic to those hosts actually passes through that edge. The cost is that any TCP client sending IP options gets cut off, and that kind of traffic is rare.

TuxCare engineers have reviewed the rule, and it looks right. With no public proof of concept to throw at it, though, there’s nothing to test it against yet.

Second, confirm the migration path is closed. sysctl net.ipv4.tcp_migrate_req should return 0, which is the default, and the host shouldn’t have a BPF_SK_REUSEPORT_SELECT_OR_MIGRATE program attached.

Third, and quickest, turn MPTCP off on any host that doesn’t need it. Run sysctl -w net.mptcp.enabled=0, and put net.mptcp.enabled = 0 in a file under /etc/sysctl.d/ so it sticks. EL 8, 9 and 10 already ship that way, so this mostly matters on distributions that keep the upstream default of 1, and on EL hosts where someone switched MPTCP on. Restart services that already opened MPTCP listeners so they come back on plain TCP. For Go services, GODEBUG=multipathtcp=0 does the same per program. It takes away the listener Anthropic’s chain needs, but it doesn’t fix the bug.

If a host is internet-facing and has no fixed kernel, start by checking net.mptcp.enabled and turning MPTCP off if nothing needs it. That’s one sysctl, it’s easy to undo, and it removes the listener the demonstrated chain depends on. The IP options rule is the next layer to evaluate. Test it in staging, confirm it drops what you expect without breaking anything, and then decide whether to roll it out. Either way, put the host first in line for a fixed kernel. If your distribution already ships a fixed kernel, install it, then reboot or apply the rebootless patch, and skip the workarounds.

Checking a Server in Under a Minute

Five commands answer most of the questions:

uname -r
sysctl net.mptcp.enabled net.ipv4.tcp_migrate_req
ss -H -lMn
nft list table ip cve_2026_98432
kcarectl --patch-info | grep 'CVE-2026-98432'

The first line shows the running kernel, but EL and Ubuntu kernels carry backports the version number won’t reveal, so compare the installed kernel package against your vendor’s advisory and the table above. A 1 for net.mptcp.enabled means applications can open MPTCP sockets, and net.ipv4.tcp_migrate_req should be 0, the default. Every line ss -H -lMn prints is an MPTCP listener. The nft line prints the table and its rules if they’re loaded, so you can confirm the drop rule is actually there, and the last line shows whether KernelCare has applied its rebootless patch for this CVE. These are quick indicators, not a full security test. For a fleet, run the same lines through whatever already handles your ad hoc commands, whether that’s Ansible, Salt or plain ssh in a loop.

Where KernelCare Stands

KernelCare’s rebootless patch for CVE-2026-98432 is out for Debian 13, in both the production and ePortal feeds. Debian 12 and Oracle Linux UEK patches are in progress. The upstream fix is small and changes no data structures, so TuxCare isn’t waiting on Red Hat. Rebootless patches for EL 8, 9 and 10 kernels on RHEL, AlmaLinux, CloudLinux, Oracle Linux RHCK and Rocky Linux are being built now, and the CVE Tracker will show them as they ship. Ubuntu patches aren’t available yet, and there’s no date to share.

Kernel KernelCare rebootless patch status
Debian 13 Released, production feed since October 26 and ePortal since October 28
Debian 12 In progress
Oracle Linux UEK In progress
EL 8, 9 and 10 (RHEL, AlmaLinux, CloudLinux, Oracle Linux RHCK, Rocky Linux) In progress, not waiting on Red Hat
Ubuntu Not available yet, no date
RHEL 7 and CentOS 7 Not affected, no rebootless patch needed

On Debian 13, hosts with automatic KernelCare updates pick the patch up on their normal schedule, or you can pull it right away with kcarectl --update. No reboot needed. Confirm it landed with kcarectl --patch-info | grep 'CVE-2026-98432'. The TuxCare CVE Tracker shows rebootless patch status per distribution as it changes.

Get KernelCare today →

The Fix Came First, and It Still Isn’t Enough

By the usual standards, this disclosure went right, and it still leaves a lot of enterprise Linux servers waiting.

Coordinated disclosure did its job upstream. The fix landed before the CVE was published, the CVE was out before anyone claimed an exploit, and the exploit code itself stayed private. Nobody can call this a zero day.

None of that helps a RHEL 9 web server with port 443 open. A two-line upstream fix still has to become a vendor kernel build, then a shipped package, then a reboot or a rebootless patch on every exposed host. For RHEL and the distributions that rebuild it, that first step hadn’t happened as of October 30. Meanwhile, a public write-up now describes the shape of a working chain and how long each stage took. The code stayed private, but the approach didn’t, and the next attempt doesn’t have to start from scratch.

So check which hosts have MPTCP listeners, especially anything written in Go, and turn MPTCP off wherever nothing needs it. Test the IP options rule in staging, and if it behaves, consider it for internet-facing servers that don’t have a fixed kernel. Then watch for fixed kernels from your distribution and work out how fast you can get them onto every exposed host – by rebootless patch or by reboot, whichever your fleet will actually tolerate.

Get KernelCare today →

How the Month Unfolded

Date Event
January 2016 Linux 4.4 ships with commit 079096f103fa, which introduced the bug
October 5, 2026 Eric Dumazet’s two-line fix is applied upstream
October 15, 2026 The kernel CVE team publishes CVE-2026-98432
October 16, 2026 Oracle fixes UEK, and Debian fixes trixie (DSA-6568-1)
October 17, 2026 Debian fixes bookworm’s linux package (DLA-4841-1)
October 22, 2026 Debian fixes bookworm’s linux-6.12 package (DLA-4850-1)
October 24, 2026 The kernel CVE team adds its 9.8 score, and Amazon fixes kernel6.18
October 26, 2026 KernelCare’s Debian 13 rebootless patch reaches the production feed
October 27, 2026 Anthropic says an unreleased Claude model wrote a remote root exploit against an MPTCP listener
October 28, 2026 KernelCare’s Debian 13 rebootless patch reaches ePortal
October 29, 2026 Red Hat moves to Important (8.1), SUSE publishes two kernel advisories, Amazon fixes kernel6.12, and the Go Security Team publishes guidance
Pending RHEL and Ubuntu kernels, SLES 15 SP7, the remaining Amazon Linux kernels, and KernelCare rebootless patches for Debian 12, Oracle UEK, RHCK-based EL 8, 9 and 10 kernels (in the build) and Ubuntu

Sources

Primary sources for this post are the kernel CVE announcement, the CVE record and the upstream fix, plus vendor pages from Red Hat (including its advisory file and bug tracker), Ubuntu, Debian, Oracle, Amazon and SUSE. Also cited are Anthropic’s post, the Go Security Team’s announcement, Go’s GODEBUG documentation, the kernel’s MPTCP documentation, the NVD entry, The Hacker News coverage, and TuxCare’s KernelCare release records as of October 30, 2026.

Join 4,500+ Linux & Open Source Professionals!

2x a month. No spam.

Table of Contents
Get the open-source security answers you need

Join 4,500+ Linux & Open Source Professionals!

2x a month. No spam.