Anthropic just made AI scarier

How powerful is AI? Enough that Anthropic, a leading AI company, announced earlier this month that its latest AI model, Claude Mythos Preview, would be available only to a limited number of businesses due to security concerns — at least for now.

Claude Mythos Preview was designed for general use, Anthropic says, but during testing, the company found it extremely effective at identifying vulnerabilities in the security systems of all types of software, creating potentially massive security concerns.

So far, Anthropic is sharing the Mythos Preview model with a handful of major tech companies and banks through a program called Project Glasswing, intended to give them an opportunity to shore up any existing security vulnerabilities and get ahead of potential hacking attempts that the model could identify.

To get a better sense of what Claude Mythos Preview represents and the potential threat it brings to online security, Today, Explained co-host Sean Rameswaram spoke with Hayden Field, senior AI reporter at The Verge.

Below is an excerpt of their conversation, edited for length and clarity. You can hear the full episode wherever you get podcasts — including Apple Podcasts, Pandora, and Spotify.

What is Claude Mythos? 

Mythos is [Anthropic’s] newest AI model that they designed to be a general-purpose AI model like any other. But what they realized when they were working on it was that it had these special skills that they didn’t really anticipate. It was really good at cybersecurity. It found high-stakes vulnerabilities in virtually every operating system.

That’s pretty bad if you are using that as a hacker. And to have a blueprint for a list of every big gap and insecurity and vulnerability on all these really, really high-profile systems, you’re going to be having a list of everything you could do to take those systems down or exploit data. 

They realized that they better not release this to the general public because it could fall into the wrong hands. And they instead handpicked a select few organizations that are responsible for critical infrastructure to release it to so they could plug those gaps in their systems instead.

You’ve heard of many of the companies that currently have and are using Claude Mythos: Nvidia, JP Morgan Chase, Google, apparently a few dozen more that build or maintain critical software infrastructure. How does it actually work? 

Since they built it as a general-purpose model, it probably works like any other model in that you’re using it and prompting it to flag all the vulnerabilities in your system.

Maybe you’re Google Chrome, and you’re looking for specific, niche parts of the browser that you think may have some vulnerabilities. You’re basically prompting the model to flag all these really high-profile gaps to you and your security, and then you’re taking that and plugging it up on your own. 

A hacker would actually use it in the same way. If it fell into the wrong hands, they’d be like, “Yeah, tell me all the vulnerabilities here.” And then they’re going to take it off the platform and use that for something nefarious. So it’s basically about who is prompting the system and what their motives are. 

It’s as easy as saying, “Hey, Claude, tell me how this banking system might be vulnerable.” And then Claude thinks about it for a minute, and it spits out a bunch of answers.

Essentially, yes.

And do we know that the Googles and Nvidias of the world are actually using this technology?

Yes. Part of the reason that Anthropic released this is they wanted these organizations to report back on exactly how Mythos worked and what it did to plug up the vulnerabilities and the gaps in their system. It’s an information-sharing thing. 

They’re letting these companies use it to test out how well it does to plug up all these high-profile gaps, and then they have to report back to Anthropic about how it worked.

How is Anthropic choosing who to share this technology with?

I actually asked them that. They’re essentially looking for cyber defenders or companies that a lot of people depend on, and that downstream it would be a huge issue if they got hacked in any way, shape, or form. 

JP Morgan Chase is a great example. Anthropic has also offered this technology to the government. 

Do Anthropic’s competitors have similar tools? Are they presumably working on similar tools?

OpenAI is apparently working on a similar tool. Anthropic itself has said this isn’t something that they deem they’ll be in the lead on for too long. They think labs anywhere in the world may release this technology in the next three months, six months, 12 months. 

It seems like, sometime in the next 12 months, this is going to be out there. And so that’s why they wanted to release Mythos now, so that companies and banks could get ahead of all the hacks that may be coming down the line, when similar types of technology are released to the general public, maybe months from now.

If this is so dangerous and there’s so many potential risks, is anyone having a conversation about just not releasing tools like this and just sort of shutting it down, keeping it internal?

That is a really great question. I’m so glad you asked, because not enough people ask whether an AI system should actually be released or used for certain things. Right now, we’re seeing a lot of one-size-fits-all, throw-it-at-everything type of integration. And a lot of times AI is not the answer for things. 

With this, though, people tend to agree that it is something that’s needed right now. AI is already out there helping cyberattackers really step up their attacks. And we’ve been seeing that intensify over the past year. People seem to agree that you need AI to fight AI cyberattacks, essentially. 

It’s kind of like medieval fortresses, where you’re adding extra stones and building up the walls at the fortress higher because a war is coming. That’s the sense I get when I talk to these experts about this. They know it’s coming. It’s just, ‘Try to shore up your defenses now so that you’re best prepared.’