No Priors: Artificial Intelligence | Technology | Startups
00:00:00 · George Kurtz, Sarah Guo
[00:00:00] S0: For the first time, security is the gas pedal, not the brake. [00:00:03] S1: How would you describe what you're seeing AI doing for adversaries? [00:00:07] S0: The new apex predator is the agent state. The same level of capability that a nation state would have just a few years ago, essentially a hacktivist would have today. That has dramatically changed the overall threat environment. What people have to keep in mind is it's not malware anymore, it's prompts. They can basically engineer prompts that would create a totally unique signature of what direction they're gonna go in, how they gather information, where they pivot. Everyone's coming to CrowdStrike and others going, well, how do you make AI secure so we can go faster in AI? The only way to fight AI today is with better AI. The defender knows their own systems. The defender has the architectural diagram. The defender has the code. How do you put all of that into something that you can dramatically increase your probability of outsmarting the offensive AI? [00:00:58] S1: Hi, listeners. Welcome back to No Priors. Today, Elad and I are here with George Kurtz. George is the world's first call when there's an incident, a legendary cyber defender, the founder and CEO of Foundstone, and then CrowdStrike. We talk about what's happening in cybersecurity, fighting AI with AI, their internal research effort, Safemind, and whether the odds are good for the Internet in the next era. Welcome back. George, thanks so much for doing this. [00:01:23] S0: It's a pleasure to be here.
00:01:23 · George Kurtz, Sarah Guo
[00:01:23] S0: It's a pleasure to be here. [00:01:25] S1: So it is a sea change period in technology, and agents are taking over the enterprise. CrowdStrike is on the forefront of that. What can you tell us about what you're you're seeing in in the wild right now? I think you have said that there are, you know, two, three magnitudes more agents out there taking action in in people's estates than they're aware of. [00:01:51] S0: It's been incredible in the enterprise what we've seen. If you follow the trajectory of the Frontier Labs and you just look at their own revenue, you can see just how many more agents are out there, whether it's Claude Code or Codex They're popping up all over over the environment. Obviously, there's many other type of agents that are out there, but again, just using that as a proxy. And you know what we've seen is that they're popping up everywhere and part of the problem in these enterprises is they have no idea where they're popping up. Shadow AI has become a huge issue for many companies. I know we're going to talk about what the agents are doing and all the bad sort of things that are happening, but just in general, having organizations understand where their AI is taking place and who's doing what, and more importantly, what identity are they actually operating under has been a big challenge in the industry and a big focus for us at CrowdStrike in trying to crack the code for our customers. [00:02:45] S1: How would you describe what you're seeing AI doing for adversaries? You've described it as flattening the adversary pyramid before.
00:02:54 · George Kurtz
[00:02:54] S0: When I think about the adversaries, I call it the pyramid of pain. So, you just imagine the bottom part of the pyramid, the base of the pyramid is really what we call a hacktivist. So you've got the least level of sophistication, but the greatest number of adversaries. Then in the middle band of the pyramid you've got e crime. They're more sophisticated and a few less than all the hacktivists. Then at the top of the pyramid you've got nation state actors, right? So the most sophisticated, but the least number relatively speaking. Well, that has been totally just blown up because really now what we're seeing is the new apex predator is the agent state, not the nation state, it's the agent state and what the agent state has done. If we think about the models that are out there that are ablated so there's no real guardrails, it's essentially taken all of these very complex topics and sort of techniques and cost and now flatten the entire pyramid. So the same level of sort of capability that a nation state would have just a few years ago. Essentially, a hacktivist would have today, and that has dramatically changed the overall threat environment. So what are we seeing? We're seeing more vulnerabilities found and we're seeing faster attacks and when the attacks come, it's just you know it's just one after another. You can see it. It's actually right now it's pretty clear whether it's an AI attack because you can see how fast they're actually iterating. But at some point they're going to introduce that being low and slow and it's going be more difficult to figure out what's AI and what isn't. But that's what we're seeing and I think that's one of the real challenges that organizations are facing is just the barrage of attacks and the level of sophistication has dramatically increased. And ultimately, if if you give an agent a task with enough time and the agent either needs to get in or dies, it's going to try to figure out a way to get in, and a lot of times it's going to have success.
00:04:50 · Elad Gil
[00:04:50] S2: Are you seeing a shift in terms of how these adversarial attacks are being made? I know we were talking earlier, we talked a little about open versus closed source models or open weight models. Are you seeing a shift where most of it is open weight models? Is it a mixture in terms of people using state of the art labs and capabilities? Like what what are people actually using to do these attacks at this point?
00:05:07 · George Kurtz
[00:05:07] S0: We're seeing a lot of it on the open weight models, right? If you look at, I mean, just the guardrails on the latest generation of any of the frontier models. To be candid, you can barely create a PowerPoint that has cyber security in it. I mean, it literally rejects you and bumps you down to a lower model. So what we've seen is, you know, again, if you look at GLM, if you look at Quinn, if you look at any other sort of models that are out there, they've been very effective in being able to deal with cyber security and many of the models are either ablated or specifically built just for cyber. It might be for red teaming and those are the type of attacks where you could actually just run these when you can run them on. I've got two spark DGXs at home and you know I've got all the models running and you know whatever I want to do I can do. You want to create ransomware? No problem. You want to create a ransomware kit? No issue. You want to deploy? You want to iterate? I mean it's all there for you and you know that is like one step removed from just running it all on your laptop. And we think about where Mac is today. Even the new Macs that are coming out, The studios they're going be incredible AI machines, so the tip of the spear is going to be what you can run without the guardrails and the frontier models are sort of helpful and in terms of if you needed PowerShell scripts or those sort of things. But certainly you can do a lot of it with the open weight models right now.
00:06:37 · Elad Gil
[00:06:37] S2: And is there a common geography? Like are most of the attacks coming from a single region? Is there a common attack type? Is that mainly ransomware? Is it something else? Is it phishing campaigns? I'm little bit curious, like has the mix shifted as you've kind of seen the adoption of AI tools?