Your AI Malware Experiments Are Showing

I mean, speaking of APT, I wanted to like ask you, Dave, like what are the kinds of things are you seeing in the state sponsored space with AI integrated malware? Are you seeing some of this stuff being used in the wild? What are you kind of tracking?

Well, it's interesting what Ryan just said, right? We see a lot of development right now and less so operationally.

So the shift to, you know, a balancing of that ratio would definitely be interesting.

And as Ryan hinted at, there's difficulties when it comes to defending that because of, you know, policies and like you can say AI integrated SOC, but when it comes to automated decision-making in a SOC environment, Policy-wise, what are you willing to live with with that? But we see it all across the attack chain, I'd say.

And it just varies how it's used depending on the type of actor.

The more sophisticated actor can use it in subtler ways.

But across the board, improved social engineering and phishing at the very least.

I mean, that feels still like the strength, the biggest strength for a skilled attacker to leverage the models, right? It's still in that space.

You know, like the core of it being so persistent and stuff also creates a lot of loud noise, right? Like, so it's kind of like, it's an interesting dynamic right now for like a more skilled attacker to leverage it in different ways.

I try to follow along and see how each model is also coming along in that space.

Not just that they score really well in terms of finding vulnerabilities or maybe starting to chain attacks better, but it's also in the how much do they create noise-wise in that process, right? Because that's a very, very different animal.

Ryan, you're tracking, is it like a dozen malware families at the moment?

It's up to like 20-something now.

Nice.

Oh, it has grown.

Okay.

What are some of the most interesting things you've found from the malware families that you currently track?

So, yeah, what I'm currently kind of digging into is like these anti-analysis strings that attackers are putting malware in their malware.

So this is where like...

the scope of Cairn includes malware that's targeting AI as well as like having AI integration into it.

And there's, I think there's 10 or 11 like archetypes that I use to classify um all of these types of malware and and the archetype is basically like what purpose is the the ai um integration serving that's how i um determine like which archetype it fits and the third archetype is basically like malware that's trying to evade ai by putting like prompt injection or instructions to the model in the binary and this is like become a really interesting space and i think one that has demonstrated the value of Cairn because as I've been preparing to write and release my next blog post, I'm seeing these posts come out about this happening in the wild.

So just the other week, there was a report from ESET about a Russian APT putting requests for helping with building a nuclear or biological weapon in one of their Visual Basics

scripts.

There was another one from Google.

I think it was called Dust Breaker, Dust Maker, which had a similar thing.