Assessing Claude Mythos Preview's cybersecurity capabilities (opens in new tab)

No, the elephant in the room is that even bad actors will now have easier to find vulnerabilities in, maintained or not, widely or in critical places used software. Unmaintained and remotely accessible devices should be discarded as soon as possible, you can't stay waiting till some of the good guys decide to give some time to your niche but critical unmaintained piece of software. Because if there is a possibility of taking profit of it, it will be checked and exploited.

And you can't assume that whatever vulnerability they have will let good guys to do the extra (and legally risky) work of closing the hole.

And this is precisely why so many of these devices should not be connected to the Internet.

Things like an Internet-connected central heating seem absolutely insane to me, yet people look at me like I'm crazy when I say so. Do you really want your home' heating entirely controller by a publicly accessible device that likely will never be upgraded in case of security issues?

Not to mention embedded systems. In fact, most people's Windows machines hardly get updated. You remember WannaCry, right? I work at a mid-sized e-commerce company making hundreds of millions in annual profit. Our servers run Windows Server 2012 and use PHP 5.3 — never upgraded. Aside from me, the newest developer machines are Windows 10 21H2, then Windows 10 1809, and even Windows 7. I heard there’s also a server running Windows Server 2008. And I don't see any hope for improvement: non-software companies, especially in the current economic climate, cannot invest huge resources to completely refactor everything. The entire tech department is no more than 10 people; doing a refactor would mean halting all business operations, so patching and mending on top of what's already there is the only viable option. Shortly after I joined, I found several SQL injection vulnerabilities and successfully exploited them to register as the root user on the server (on MySQL 5.5) and extract passwords. This is the technical reality for many non-specialist software companies.

You should either implement over-the-air updates or not connect your device to the network at all.

> The only practical defense is for these frontier models

Another practical defence for many of these devices would be to just disconnect them... I feel like an old man yelling at a cloud, but too much is connected to the Internet these days.

The majority of vulnerabilities are in newly committed lines of code. This has been shown again and again [1] [2]

From a marketing standpoint Anthropic is showing that they're able to direct 'compute' to find vulnerabilities where human time/cost is not efficient or effective.

Project Glasswing is attempting to pay off as many of these old vulnerabilities as possible now so the low-hanging fruit has already been picked.

The next generation of Mythos and real world vulnerabilities exploits are going to be in newly committed code...

[1]: https://dl.acm.org/doi/epdf/10.1145/2635868.2635880

[2]: https://arxiv.org/abs/2601.22196

I'd love for them to target their own code base considering we keep seeing security vulnerabilities in claude code.

How likely is it that they're not using their latest and greatest for their own projects though? Perhaps their ability to find security flaws is surpassed by their ability to create them.

> Mythos Preview identified a memory-corruption vulnerability in a production memory-safe VMM. This vulnerability has not been patched, so we neither name the project nor discuss details of the exploit.

Good morning Sir.

> Has anything changed here? I don't pay much attention but KASLR was considered basically useless for preventing LPE a few years ago.

No. It's still like this. Bonus point that there are always free KASLR leaks (prefetch side-channels).

But then, this thing is just.. I don't have a word for this. Just randomly read paragraphs from the post and it's like, what?

I love the goal post shifting. All modern code is ai slop right? Isn't the whole point we hate ai bc it generates vulnerable slop?

Nope, not allowed to attack bsd code, it's gotta be electron-shit #9001 or we can't trust it

There is a lot to digest here. Maybe having a few separate pages makes them a bit more digestible. The system card itself is some 200 odd pages

Would vote to keep them separate. They seem independent enough to warrant their own discussion based only (or rather, mostly) on the content from each link. edit: merging this and glasswing as underdeserver stated would probably be fine

I think the system card one should be separate, but this and the Glasswing thread are basically the same story.

Construction is always more expensive than destruction

offense has a clear reward function, but so does detection when you frame it right. "did this process try to read ~/.ssh/id_rsa?" is just as binary as "did the exploit land?" the reason defense feels harder is that people frame it as architecture review (fuzzy, subjective) instead of policy enforcement (binary, automatable). we keep trying to make AI understand intent when we should be writing rules about actions. a confused deputy from 1988 doesn't care why the request came in, it cares whether the caller is authorized. same principle applies here.

A plateau is unlikely, at least for cybersecurity. RL scales well here and is replicable outside of Anthropic (rewards are verifiable, so setting up the training environment doesn't require that much cleverness).

The post also points out that the model wasn't trained specifically on cybersecurity, and that it was just a side-effect – so I think there's still a lot of headroom.

It's scary, but there's also some room for cautious non-pessimism. More people than ever can cause billions of dollars of damage in attacks now [1], but the same tools can be used for defensive use. For that reason, I'm more optimistic about mitigations in security vs. other risk areas like biosecurity.

[1]: https://www.noahlebovic.com/testing-an-autonomous-hacker/

On a topic like cybersecurity, we never win by not looking: One needs top of the line knowledge of how to break a system to be able to protect it. We have that dilemma dealing with human experts: The same government sponsored unit that tells you that you need to update your encryption can hold on to the information and use it to exploit it at their leisure.

Given that it's absolutely impossible to stop people not aligned with us (for any definition of us) from doing AI research, the most reasonable way forward is to dedicate compute resources to the frontier, and to automatically send reasonable disclosures to major projects. It could in itself be a pretty reasonable product. Just like you pay for dubious security scans and publish that you are making them, an LLM company could offer actually expensive security reviews with a preview model, and charge accordingly.

We need to promote alignment and other ethics benchmarks; we can't change what we don't measure. I don't even know any off the top of my head.

The immediate plateau is the energy output of the Sun captured by the Dyson Swarm around it. Until there it's smooth sailing.

If we don't innovate, someone else will. This is the very nature of being a human being. We summit mountains, regardless of the danger or challenge.

It would certainly be good news for cybersecurity employment!

Gives me Fight Club vibes.

Well, maybe not... see Simon Willison's ongoing reporting [0] on all the bug reports for `curl` people are finding with LLMs.

Interesting to see them go from "DON'T GIVE US AI SLOP!" to "Wow, lots of actual bugs found, including [ed: at least one] bug found by two people!"

[0]: https://simonwillison.net/search/?q=curl

Keep in mind that Opus detected most of these vulnerabilities, it just didn’t exploit them (says so much in the article).

I’m honestly not convinced this is changing the landscape significantly. It’s simple a bit better at self directing.

The name also reminds me of Final Fantasy XVI when Ultima was calling Clive, Mythos, as Meee-thos