|
|
Log in / Subscribe / Register

Remote code execution on Linux systems with multipath TCP

On October 27, LLM provider Anthropic reported on a remotely exploitable use-after-free bug affecting systems with multipath TCP listeners open. The bug works by triggering a response to a retransmitted SYN packet at the same moment that the TCP connection is established in another thread, causing the first thread to access the connection's options structure after the second thread has freed it. The first thread can then disclose kernel memory or cause a buffer overflow, writing an attacker-controlled payload. The fix was to only free the structure after an RCU grace period. On October 29, the Go security team advised that all programs written in Go use multipath TCP by default, so many systems may be unknowingly vulnerable. And, of course, there is always the possibility of related bugs in other sections of the TCP stack. Users may be wise to update to a patched kernel as soon as one becomes available for their distribution.



to post comments

Debian fixed it some days ago

Posted Oct 30, 2026 16:32 UTC (Fri) by storner (subscriber, #119) [Link]

Thank you

Posted Oct 30, 2026 18:08 UTC (Fri) by chris_se (subscriber, #99706) [Link]

Thank you very much for the important information that Go uses multipath by default - because otherwise I'd have assumed that nothing I'm running uses MP.

So... hmm... Kubernetes?

Posted Oct 30, 2026 18:19 UTC (Fri) by NightMonkey (subscriber, #23051) [Link] (7 responses)

Go uses Multipath TCP by default. There's *lots* of stuff that is written in Go that offers network services over TCP... including big ones like Kubernetes. Is there any centralized resource to see at a glance which network-facing projects are vulnerable and which version has a fix? Thanks.

So... hmm... Kubernetes?

Posted Oct 30, 2026 19:08 UTC (Fri) by nicholascapo (subscriber, #131674) [Link]

The vulnerability is in the kernel, not in Go, Go just happens to be a (common) user of MPTCP.
If your kernel isn't up to date, then anything (bash scripts?) using MPTCP is vulnerable.

So... hmm... Kubernetes?

Posted Oct 30, 2026 19:10 UTC (Fri) by Cyberax (✭ supporter ✭, #52523) [Link] (5 responses)

It's basically everything that has a listener. You need to either update the kernel or turn off the MPTCP: sysctl -w net.mptcp.enabled=0

So... hmm... Kubernetes?

Posted Oct 30, 2026 21:43 UTC (Fri) by dskoll (subscriber, #1630) [Link] (4 responses)

I compile my own kernel packages and the sysctl command you gave failed. So I checked:

$ grep -i mptcp /boot/config-7.3.1 
# CONFIG_MPTCP is not set

I guess I'm accidentally safe.

So... hmm... Kubernetes?

Posted Oct 31, 2026 1:22 UTC (Sat) by job (guest, #670) [Link] (3 responses)

MPTCP is just the only publicized exploitation path, there could be others.

So... hmm... Kubernetes?

Posted Oct 31, 2026 18:29 UTC (Sat) by dskoll (subscriber, #1630) [Link] (2 responses)

Fixed in 7.3.1, though?

So... hmm... Kubernetes?

Posted Oct 31, 2026 19:40 UTC (Sat) by job (guest, #670) [Link] (1 responses)

Should be. Although it will be very interesting to know why Red Hat, Canonical and SuSE all haven't released their kernels yet. It's basically a one-liner. Either it brings some non-obvious regression, or they suspect there is more to the fix than this.

So... hmm... Kubernetes?

Posted Oct 31, 2026 21:11 UTC (Sat) by gutschke (subscriber, #27910) [Link]

Proxmox claims that they got the fix worth the upstream Ubuntu 7.0.0-43 kernel. So, maybe it's already fixed and just not documented as such?

This is bad

Posted Oct 30, 2026 21:12 UTC (Fri) by Cyberax (✭ supporter ✭, #52523) [Link] (1 responses)

This is bad. I was able to build a POC crash generator in about an hour with Claude. Presumably, a full RCE is only a bit more tricky.

I've updated everything I could, and set a rule on my firewalls to drop any TCP SYN packets that have any IPv4 options enabled. And I'm already seeing this rule activating.

This is bad

Posted Oct 31, 2026 19:52 UTC (Sat) by job (guest, #670) [Link]

"A bit" might be underselling the complexity. It would be very impressive if anyone could pull that off outside a controlled system. But yes, same here and went with:

iptables -I INPUT 1 -m u32 --u32 "0 & 0x0F000000 >> 24 = 6:0xF" -j DROP

Not seeing any hits though, I suspect my upstream router refuses to route such packets. I believe that is part of the reason why things like multicast had a hard time on the public Internet.

Fixed in LTS releases from 2026-10-15 - 7.2.11, 6.18.57, 6.12.113, 6.6.159, 6.1.190, 5.15.223, and 5.10.272

Posted Oct 30, 2026 22:03 UTC (Fri) by himi (subscriber, #340) [Link]

. . . at least for the newer versions - I can say for sure about the 6.12.y branch, and given it goes back to the 4.4 kernel I'd be surprised if it hasn't been backported to all the active LTS releases.

The backported fix in the 6.12.y branch is:

commit 11a03ce3c4a62d84c10ccdfe47f8075006498568
Author: Eric Dumazet <edumazet@kernel.org>
Date:   Thu Oct 1 22:12:53 2026 +0000

    ipv4: free inet_opt and ireq_opt after an RCU grace period
    
    commit e3f33b0a1d89d738e2b913dab519226eaf76e15b upstream.

The RCE is the scary thing here, particularly given Go's behaviour.

One positive from the last six months of kernel hell is that at my organisation we're proactively following the 6.12 LTS releases, with no more than a two week lag on deployment (unless there's a serious known vulnerability in play, which was actually the case for 6.12.113) - that means when this RCE was announced I could sit back and enjoy the fact that my externally exposed systems had been safe for several weeks.

Linux needs a sysctl to disable IP options

Posted Oct 31, 2026 20:32 UTC (Sat) by rbranco (subscriber, #129813) [Link] (3 responses)

OpenBSD PF blocks packets with IP options set by default.

https://www.openbsd.org/faq/pf/filter.html#ipopts

FreeBSD has net.inet.ip.process_options=0

Any chance we can add this sysctl to Linux?

Linux needs a sysctl to disable IP options

Posted Oct 31, 2026 23:00 UTC (Sat) by brianbents (guest, #130799) [Link] (2 responses)

MPTCP is encoded with "TCP Options" though, not "IP Options" ( https://en.wikipedia.org/wiki/Multipath_TCP#Multipath_TCP... )

Linux needs a sysctl to disable IP options

Posted Nov 1, 2026 0:36 UTC (Sun) by brianbents (guest, #130799) [Link]

Oh, and there is a sysctl explicitly for MPTCP, "net.mptcp.enabled" (as well as others to tune it, if using).

Linux needs a sysctl to disable IP options

Posted Nov 1, 2026 1:02 UTC (Sun) by Cyberax (✭ supporter ✭, #52523) [Link]

This bug is triggered by IP options processing, MPTCP allows it to happen in parallel.

I've been using MPTCP quite happily at home. This is unfortunate.

Posted Nov 1, 2026 20:05 UTC (Sun) by ejr (subscriber, #51652) [Link] (4 responses)

I can unplug a laptop and ssh sessions stay. Remote displays stay.

Also, Claude Code is the agent, not the model. Agents more tuned towards security likely would find even more problems.

I've been using MPTCP quite happily at home. This is unfortunate.

Posted Nov 2, 2026 5:12 UTC (Mon) by lindi (subscriber, #53135) [Link] (3 responses)

Is there a solution that would make this work over the Internet or is this realistically only going to work in your own network where you can make sure nothing blocks MTCP?

I've been using MPTCP quite happily at home. This is unfortunate.

Posted Nov 2, 2026 15:54 UTC (Mon) by ejr (subscriber, #51652) [Link] (2 responses)

It does work, but not from everywhere. Works at a coffee shop I frequent, but I doubt if it'd work on a corporate network.

I've been using MPTCP quite happily at home. This is unfortunate.

Posted Nov 2, 2026 18:34 UTC (Mon) by Cyberax (✭ supporter ✭, #52523) [Link] (1 responses)

It works in most networks that don't take active measures to prevent it. MPTCP looks like regular TCP for firewalls.

I've been using MPTCP quite happily at home. This is unfortunate.

Posted Nov 2, 2026 19:26 UTC (Mon) by ejr (subscriber, #51652) [Link]

A bunch of firewalls clear all option bits. Certainly many corporate ones I've encountered do that... Not a good idea, but also not my choice.


Copyright © 2026, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds