CVE-2026-98432: Linux TCP Use-After-Free Risk and Patch Validation | Hacker News Reader

CVE-2026-98432: Linux TCP Use-After-Free Risk and Patch Validation

penligent.ai

65 pointscobychapple43 comments

Whoever failed to edit this CVE description down owes us all a bunch of our time back

Everyone using LLMs needs to stop shipping their first draft.

"Haven't had enough yet of very junior engineers pasting wall-of-text from chatgpt at you in Slack? Come on down to the CVE announcement emporium, where we have a special treat for you today..."

What about senior engineers doing that?

And I'm the junior engineer getting annoyed at the senior engineers at my company doing it...

You know it isn't serious when it doesn't have its own domain name.

Blaise Pascal: December 1656

"I have made this letter longer than usual, only because I have not had the time to make it shorter"

("Je n’ai fait celle-ci plus longue que parce que je n’ai pas eu le loisir de la faire plus courte").

> Blaise Pascal: December 1656

Also Balzac:

* https://quoteinvestigator.com/2012/04/28/shorter-letter/

Thankfully the vast majority of servers are not creating MPTCP listeners by default (that I'm aware of).

Not sure why the article doesn't mention it, but you can specifically check for MPTCP listeners using `ss -M`.

https://go.dev/doc/go1.24

> ListenConfig now uses MPTCP by default on systems where it is supported (currently on Linux only).

This seems to imply basically any golang server built since February 2025 is going to be at least attempting to use it.

Yeah seems you're right...

Luckily though, ListenConfig lets you override this https://cs.opensource.google/go/go/+/refs/tags/go1.27.3:src/...

mptpcp usage is not a prerequisite

Do you have a source for this claim?

Everyone who’s read the patch.

Want a big name? RedHat.

Is there a PoC that exploits this without MPTCP?

You need `ss -Ml` if your server isn't actively serving MPTCP clients.

I was vulnerable myself by using Caddy (although MPTCP is not a prerequisite).

Would dropping these specialized ipv4 syn packets with extra header size beyond 20 bytes (as indicated by options value >5) with the firewall be low enough in the order of things to mitigate the attack? Like, iptables -A INPUT -p tcp --syn -m u32 --u32 "0>>24&0xF=6:15" -j DROP ?

I fail to understand / find the gist of this article.

Article describes the technical aspects of the bug, how it got handled by different vendors...AI position on it...

But, one would always ask themselves, why did the author write this article? That part hasn't been clear to me even after going through much of the article.

I believe the author of this article is in fact most likely a large matrix multiplication and that it is possible that no human even initiated the computation of said multiplication directly. I assume it's not hard to set up some monitor for new CVEs to try and slop out an article as quickly as possible to get as revenue

Because it’s slop.

The overlaid HTML popup on that page for "Try The World's First Agentic AI Hacker." tool is cancerous and annoying. This should really be a link to a more reputable link for the CVE in question.

It doesn't appear until you've read the page for maybe 10-15 seconds, so if you're not seeing it, just wait.

It even managed to get past ublock origin, since there's no specific rule for it yet.

Yes I found it very jarring too

This is a good portion just slop, another 9000 word article that could have been a couple of paragraphs. More text doesn't mean better.

There's other coverage, this just has most of the comments now

Ubuntu (34 points, 16 hours ago, 3 comments) https://news.ycombinator.com/item?id=50287647

CVE.org (2 points, 7 hours ago) https://news.ycombinator.com/item?id=50295879

GoogleGroups (slight tangent) Go programs enable MPTCP by default, exposing them to Linux CVE-2026-98432 (9 points, 3 hours ago, 4 comments) https://news.ycombinator.com/item?id=50298526

It would have been about 50x less annoying if whoever was driving the LLM had simply told it to make a clickable top of page index/headers in markdown and divide it into sections and subsections by logical topic.

Given these types of security bugs that are being found left and right, I reckon there's going to be an uptick in hiring of software engineers to fix them.

I strongly doubt that. The fixes for most of these are trivial to the point it’s almost automated.

There is a push for that, but the fixes (often as little as a one line Dockerfile change in some cases) still have to be tested and verified by a real engineer.

Otherwise everywhere would have patched already with AI.

The resulting mechanical code change might be trivial. The bugs themselves are often decidedly not.

You're essentially saying that swallowing a pill is trivial, so therefore pharmaceutical engineering is.

For most of these bugs, they occur in a 3rd party software thats easily upgraded.

I’m saying patching is easy because the pill is delivered to you.

I suspect a lot of organisations are already on automated AI detect-patch-redeploy loops for a lot of these types of issues.

This one is quite serious and the type of vulnerability is once again, a C style memory corruption and a race condition combined; leading to an RCE.

At this point, we seriously need to discuss whether if the major projects still written in C such as Linux, Chromium and even git can even handle the stream of now near infinite bugs being discovered with AI.

Attackers are already rapidly finding and building exploits for those vulnerabilities the moment they are discovered with abliterated frontier-level models.

These sort of vulnerabilities are far to common in those projects and in the future, it will be extremely unsafe for even experts to continue using C or C++ in any way.

Leading 'maybe' to an RCE, i imagine you need to groom the heap first, unless there is something i'm missing.

Edit: also thanks for using RCE correctly, it seems to be any kind of code execution these days is considered "RCE".. boo.

I bet there's a half-life to bugs though. Bugs might be infinite over the lifespan of a project, but the rate that they are produced goes down with our ability to detect them.

A lot are being found _now_ in projects written in C because the backlog of bugs to find is mostly written in C. We haven't yet gotten to the point when we're detecting most bugs before they're released, but we will.

Things won't get worse for too long. If we haven't already surpassed it, we're approaching the point of being able to do static analysis and ensure the mathematical correctness of the most complex C programs ever written so far.

I don't use C or C++, but look how much safer it is to use C or C++ these days.

Unless someone vibecodes a new OS and convinces everyone to use it, we're stuck with what we got. Reaching feature parity with Linux and then debugging it to the point where it'd be usable is a gargantuan undertaking.

Don't tempt DHH. Omarchy works WAY better than expected, and made my 2017 MBP very usable again.

It just uses the same linux kernel as everyone else. So I don't see what the point of your comment is.

DHH and team are writing Omarchy components in Rust when possible. I was playing on that in my now downvoted comment.

Obviously Omarchy's kernel, Wayland, and the Hyprland compositor are not Rust and work just fine. If you've seen any of DDH's public speaking, he's willing to burn tokens to Rust all the things, don't tempt him! That was the joke, people.

I am pleasantly surprised to be daily driving Omarchy with minimal friction, by the way. For a relatively new Linux OS, this is no small feat.

You severely underestimate humanity's goalpost moving capabilities.

The state of cope:

- models are still not intelligent, they're just like, uhh, really good semantic fuzzers! so you can just put them in the loop, they'll fuzz your code, and all is good :)

- the number of bugs is "finite", so - esp with agentic help - we'll just catch up and all will be good :)

Annoyingly, the cope frontier is kinda right about both of these, so the joke shall go on. The only real counter to it would be to bring up that this finiteness claim only really applies to a frozen set of security boundaries and attack surfaces, but they'll probably just raise this being "an academic tier concern" and just tell you to ""fuzz"" more.

So ironically, it looks like these practices are rescued once again, and the bare minimum stays winning. It also starts being moved to the territory of the magic box, but I guess nobody cares about that anymore. Local AI will just get really good too, fully competitive and self hostable even. Any day now.

From the horse's mouth, in case you were suspecting these descriptions to be in bad faith: https://www.youtube.com/watch?v=NnV_cWeoo5Q

> Attackers are already rapidly finding and building exploits for those vulnerabilities the moment they are discovered with abliterated frontier-level models.

Are they? Distributions are rating this at a lower severity presumably because it can’t be easily exploited.

There's no end to bugs like these in Linux. Millions of Lines of Code can easily hide many, many of these. Code that runs in supervisor mode.

None of this can ever affect seL4[0] nor a system built on it, due to its architecture.

0. https://sel4.systems/

Year of the sel4 desktop, when?

Reply on news.ycombinator.com