22 Oct 2026 · 45m · George Kurtz, Sarah Guo, Elad Gil · 23 quoted facts
George Kurtz, CrowdStrike's CEO, argues that AI agents have flattened the pyramid of attackers, so low-skill hackers now have abilities once reserved for nation states. Attacks run at machine speed, and the only workable defense is better AI for defenders. CrowdStrike is building its own security models, with Nvidia, because mainstream AI labs' safety refusals get in the way of defensive work.
The central tension is that offense moves faster than defense, and defenders lack comparable AI tools. Kurtz uses the Hugging Face incident to show defenders turning to open-weight models after frontier labs refused to help. He is still optimistic. He calls the doom talk overblown, expects defense to keep pace, and sees security as the thing that lets businesses adopt AI faster.
Kurtz says unrestricted models give hacktivists the capability a nation state had a few years ago. The result is more vulnerabilities found and faster, relentless attacks.
Kurtz says frontier models refuse even a cybersecurity PowerPoint, while open-weight models such as GLM and Qwen can be run at home and stripped of limits. That lets attackers build ransomware kits and iterate freely.
CrowdStrike tracks how long an intruder takes to move from the first compromised machine to another. The average is 29 minutes and the fastest seen is 27 seconds, which Kurtz ties to the speed of AI inference.
In the Hugging Face incident, the company asked frontier models what was happening and was refused. It had to analyse the data with open-weight models, which is why CrowdStrike built its own red and blue security models.
Kurtz says a general remediation task costs about $10 on a top frontier cyber model, versus about $0.03 on CrowdStrike's. He credits post-training on over a decade of proprietary security data and a purpose-built harness.
Kurtz describes attackers as a pyramid: many unsophisticated hacktivists at the base, fewer criminals in the middle, and a few nation states on top. AI agents give the base the abilities of the top, so even low-skill attackers can now do advanced attacks.
Kurtz says the newest frontier models have strict guardrails and refuse security-related requests. Open-weight models, whose trained settings anyone can download and run, can be run locally and modified to remove those limits.
Kurtz says CrowdStrike measures 'breakout time', how long an intruder needs to jump from the first machine to another one. It averages 29 minutes and has been as low as 27 seconds.
SafeMind is CrowdStrike's own security AI system, with a red (attacking) model called Tempest and a blue (defending) model called Solano. Kurtz says it was built, with Nvidia, so defenders get frontier-level AI without the refusals of mainstream models.
No. He calls the idea of shutting down the internet nonsense, saying agents do what humans do but faster and have not invented new attack techniques. AI creates more vulnerabilities and faster attacks, but defenders can keep up with better AI.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
George Kurtz argues AI agents have given hacktivists capabilities that only nation-state attackers had a few years earlier.
“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.”
CrowdStrike reports average attacker breakout time of 29 minutes, with the fastest observed at 27 seconds.
“On on average, it's twenty nine minutes, but we've seen it as quick as twenty seven seconds.”
CrowdStrike's platform has grown from one module at its founding to 33 modules.
“We've got 33 modules today, and, you know, ultimately, when I started the company, we had one module, and now we've got 33.”
Kurtz says CrowdStrike pioneered endpoint detection and response (EDR), which in its early days relied on a computer and a human analyst.
“So we pioneered EDR endpoint detection response, which a lot of it in the early days was a computer and a human.”
CrowdStrike's SafeMind system includes a red model named Tempest and a blue model named Solano.
“we've got Red, Tempest and Blue Solano, those are the two models.”
CrowdStrike used Nvidia's Nemotron model as one of the main bases for its blue defensive model.
“NemoTron is one of the main models that we leveraged on on the blue side.”
At Nvidia CEO Jensen Huang's urging, CrowdStrike built a digital twin that recreates a customer's environment virtually so red and blue models can be iterated against each other.
“So we built a digital twin where we can take all of the information that we know about a customer's environment.”
Kurtz says a general remediation task on a top frontier cyber model costs about $10, compared with about $0.03 on CrowdStrike's models.
“for a general remediation, if you use like, uh, uh, uh, the highest caliber frontier cyber model, probably cost you about ten.”
Kurtz says CrowdStrike has held customer security data for about 15 years, which it uses to post-train its models.
“We've been entrusted with that for the last 15 years.”
Kurtz says CrowdStrike's endpoint sensors generate roughly 7 trillion events per day.
“So we're creating 7 trillion or so events like, uh, a day.”
Kurtz says the main lesson of the Hugging Face incident was that defenders lacked AI capabilities matching the attackers'.
“My biggest takeaway was that the defenders didn't have the same capabilities as the offense.”
Kurtz says Hugging Face turned to open-weight models to analyse the incident data after frontier models refused its requests.
“They actually had to go to open weight models to kinda sift through all this data and figure it out.”
Kurtz says the latest frontier models heavily restrict cybersecurity requests, even rejecting a PowerPoint on the topic.
“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.”
Kurtz says North Korean attackers are very active, driven by the need to bring money into their economy, often through ransomware.
“They've got to, shall we say, get money into their economy, so they've been very busy on, uh, you know, stealing information.”
Kurtz says CrowdStrike and Nvidia have worked on the SafeMind models together for about six months.
“We've been working on it for the last six months, and we kinda went through how this would work.”
Kurtz predicts the operating system of AI will be open source and open weight, as Linux is for the internet.
“the operating system of the internet is Linux, like, period, okay, and there's a reason for that, and there's gonna be an operating system of AI”
Kurtz says a large New York bank CIO told him AI costs had spiraled and that heavy token use stopped once large bills arrived.
“The costs were out of control, you know, token maxing, you know, stopped when people got big bills.”
A bank CIO told Kurtz that open-weight models, even three months behind the frontier, are good enough for most of the bank's needs.
“the models are so the open weight models right now are so good, even if they're three months behind, it's plenty for what we need.”
Kurtz says it will be hard to stop distillation of frontier models by Chinese labs because much has already occurred.
“I think the challenge is the the genie's out of the bottle. I mean, there there's a lot of distillation already taken place.”
Kurtz says breaches would persist even with no software vulnerabilities because many attackers log in with stolen identities instead of breaking in.
“What people don't necessarily realize, uh, is that even with with zero vulnerabilities, which I hope the industry can get there, you're still gonna have breaches.”
Kurtz says CrowdStrike displaced older security incumbents such as Symantec and McAfee.
“we disrupted the incumbents, we disrupted the semantics, the MacAfee's, you know, all those other players are, like, all gone for the most part, or changed or bought or whatever.”
Kurtz says he was personally involved in designing CrowdStrike's Falcon Guardian product through small-team, frequent meetings.
“we've done that with, like, our Falcon Guardian product. And to your point, like, I was personally involved, you know, had three day a week meetings”
Kurtz dismisses the idea that the internet must be turned off or air gapped because of AI agents, calling it nonsense while acknowledging the real risks.
“I think it's nonsense, you know, uh, it like, is security and and what agents can do and and the risk that they pose is that serious?”
Companies have far more AI agents running than they know about. Knowing where they are and which identity they use is a major security problem.
“Shadow AI has become a huge issue for many companies.”
Kurtz explains how unrestricted AI gives weaker attackers advanced capabilities and makes attacks faster and more relentless.
“the new Apex Predator is the agent state, not the nation state, it's the agent state.”
Attackers favor open-weight models without guardrails, and intrusions now spread across a network in minutes or even seconds.
“On on average, it's twenty nine minutes, but we've seen it as quick as twenty seven seconds.”
The incident showed defenders lacking AI as capable as the offense, because frontier models refused to help them analyse it.
“My biggest takeaway was that the defenders didn't have the same capabilities as the offense.”
CrowdStrike and Nvidia built red and blue models trained against each other in a digital copy of a customer's environment.
“you have to have a digital twin. If you look at other industries, they have digital twins.”
Kurtz argues his models are cheaper, trained on unique proprietary data, and keep customer data private.
“there is no Reddit of CrowdStrike security data.”
Large enterprises are moving to cheaper open-weight models, and the discussion covers whether distillation can be stopped and who will lead open models.
“the open weight models right now are so good, even if they're three months behind, it's plenty”
Kurtz rejects calls to air gap the internet, makes the case that defense can win with better AI, and argues technology shifts create more jobs.
“The only way to fight AI today is with better AI.”
For the 1st time, security is the gas pedal, not the brake.
How would you describe what you're seeing AI doing for adversaries?
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 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?
Hi listeners, welcome back to No Priors. Today, Elad and I are here with George Kurtz. George is the world's 1st 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, Safe Mind, and whether the odds are good for the internet in the next era. Welcome back, George. Thanks so much for doing this.
Uh, it's a pleasure to be here.
So, uh, it is a sea change period in technology, and agents are taking over the enterprise. Uh, CrowdStrike is on the forefront of that. Um, what can you tell us about what you're, you're seeing in in the wild right now? Uh, I think you have said that, uh, there are, you know, 2, 3 magnitudes more agents out there taking action in in in people's um estates than they're aware of.
It's been incredible in the enterprise what we've seen. If you follow the trajectory of the 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. Uh, they're popping up all over the, uh, all over the environment. Obviously, there's many other type of agents that are out there, but again, just using that as a proxy. And, um, you know, what we've seen is that they're they're popping up everywhere, and part of the problem in in these enterprises is they have no idea where they're popping up.
Shadow AI has become a huge issue for many companies. Um, I know we're gonna 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, uh, has been a big challenge in the industry and, uh, a big focus for us at CrowdStrike in in trying to crack the code for our customers.
How would you describe what, uh, you're seeing AI doing for adversaries? You've described it as flattening the the adversary pyramid before.
When I think about the the adversaries, I I call it the pyramid of pain. So if you just imagine the bottom part of the pyramid, the base of the pyramid is really what we call the hacktivists. 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, um, and a few less than all the hacktivists. And 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 the models that are out there, um, that are bladed, so there's there's no real guardrails. It's essentially taken all of these very complex topics and sort of techniques and cost and now flattened 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, um, because you can see how fast they're actually iterating.
But at some point, they're gonna introduce that being low and slow, and it's gonna 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, you know, if you give an agent a task with enough time and, you know, the agent needs either needs to get in or it dies.
It's gonna try to figure out a way to get in, and a lot of times it's gonna have success.
Are you seeing a shift in terms of how these adversarial attacks are being made? I know we were talking earlier, um, we talked a little bit 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?
We're we're seeing a lot of it on the open weight models, right? If you look at, I mean, just the guardrails on on the latest generation of any of the frontier models, um, to be candid, you, c- 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 if you look at, um, GLM, if you look at, you know, uh, Quinn, if you look at any of the sort of models that are out there, uh, they've been very effective in being able to, um, deal with cybersecurity.
And many of the the models are either ablated or t specifically built just for cyber. It might be for red teaming and and those are the type of attacks where you can actually just run these. I mean, 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 wanna do, I can do. You wanna create ransomware, no problem. You wanna create a ransomware kit, no issue.
Uh, you wanna deploy, you wanna 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 your laptop. And we think about where Mac is today, even the new Macs, uh, that are coming out, right? The studios, they're gonna be incredible AI machines. So the tip of the spear is gonna be what you can run without the guardrails. And the frontier models are sort of helpful in 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.
And is there a common geog- geography, like, are most of the attacks coming from a single region? Um, are the is there a common attack type? Is it mainly ransomware? Is it something else? Is it phishing campaigns? I'm a little bit curious, like, has the mix shifted as you've kind of seen the adoption of AI tools?
We sort of look at it, maybe just reframe the question a little bit, because we look at it from an adversary perspective. You know, is it a nation state? Is it e-crime? Nation state? We sort of, you know, we've come up with these names of bears and pandas, depending if they're Russia or, you know, could be from Iran, kittens and other things. So, uh, but we've seen a lot of attacks, uh, we see North Korea being very busy, uh, with these sort of attacks.
Uh, they they've got to, shall we say, get money into their economy, so they've been very busy on, uh, you know, stealing information. Uh, a lot of it around ransomware and getting paid. But, um, yeah, I think overall what we've seen is a lot of, uh, we've seen obviously out of China, a lot of activity using AI, and we've also seen it in the e-crime world where, um, when someone gets onto a system, we track breakout time.
That is, if, if, if they got onto your system, how long would it take them to actually break out and pivot and move to a different system? Uh, generally, the system they first land on is not the most interesting it could be, but it's not the most interesting. The crown jewels might be somewhere else. So on A that used to be days and, you know, hours and now it's minutes. On on average, it's twenty nine minutes, but we've seen it as quick as twenty seven seconds.
And essentially, you know, now what we're looking at is the speed of inference. So as soon as they get onto the machine, you can just just see it r- iterating very quickly. And what people have to keep in mind is it it's not malware anymore, it's prompts. So they can basically engineer prompts that would uniquely create create a totally unique signature of what direction they're gonna go in, how they gather information, where they pivot, um, which makes it, you know, more challenging to figure all that out.
But that that's what we're seeing right now, and it it's across the board from the nation state and e-crime actors, for sure.
You have been, um, a leader in security since, you know, Foundstone, CrowdStri- the founding of CrowdStrike. Um, w- would you say that there, like, how do you put the current period of, um, increase in capability, like attacker capability and perhaps defender capability, in context? Right? Are we in a more adversary dominant world right now? Like, h- how do you think about the problem overall?
Well, I think it's one of the most exciting times to be in technology. I mean, really, I've been doing this 30 plus years. And, you know, what I always like to say is that security parallels the slope of the technology curve. When I 1st got into security, it was a Cisco router and, like, maybe a website, barely right. There, there really weren't firewalls. There were these toolkits, TIS toolkit and these other things, probably dating myself. But, um, you know, there wasn't much there.
And now, if you think about the complexity of where we are today with with cloud, with, um, all of the other, uh, technologies around AI, I'm not gonna go through all of them, but it's a lot more complex today than it was way back then. So you need all these little companies and including big companies that are able to secure all these technologies. So when we think about where we are today, it actually is the most exciting, but it also creates the biggest opportunity.
Because the adversaries generally find the gaps or the seams, so they'll find new technologies that maybe that they don't have good tech the protection, um, that they're taking advantage of things like NPM packages, right? In the old world, a as you as you very well know, was no one in corporate America would just download anything and run it on their computer. It had to be vetted and scanned and all these sort of things. And now we unleash an agent, we give it an identity, which is essentially our own or worse, and then we unleash it.
And and we say, do whatever you want to do. And by the way, download whatever package you want to download. And obviously, these are the kind of things that cause a tremendous amount of exposure. So we're we're sort of getting to grips
And nobody's reading any of the production code as it gets
no, they're not. I mean, y- you know, you, you see some of these things in the wild and y- you're scratching your head. You shouldn't be surprised. I mean, it was like, let's give it a task and we'll, we'll come back in a few days and we'll hope it all worked. Nobody's watching it, nobody understands what it's doing, uh, and it's problematic. So, you know, we can, we can talk about sort of the, uh, explainability of what these agents are doing, which is gonna be really important.
Um, but it it right now, it's like a black box and nobody knows what's going on. They hope, you know, something happens and pops out on the on the other side of it. But it's, it's, it's a great time to be in technology. AI is fundamentally gonna change the world, and the frontier labs are doing amazing things. But ultimately, it it's gonna come down to security, enabling these sort of technologies. And for the 1st time, security is the gas pedal, not the brake.
And that's what we're seeing.
What is the moment that you realized this inflection was happening, at least in the security world? And I think for a lot of people, kind of the Hugging Face open AI issue was kind of like a big, you know, moment where people kind of realized there's a big sea shift in terms of capabilities and technology and what could happen. And to your point, or things being monitored, et cetera, was was that a seminal moment for you, or what are some of the seminal moments?
Let me ask you a question 1st. So what was your big takeaway from Hugging Face?
that anybody running frontier models needs to, um, uh, be careful of their environment quality and have very good ongoing production monitoring.
Yeah, and that is the takeaway. But I think what a lot of people would say is, well, this was the the 1st agentic end-to-end attack. All true, and you kinda look at, you know, what these things were doing, and, you know, people, uh, talk about them, you know, hiding themselves and, you know, trying to basically survive to the next generation and those sort of things. All true. My biggest takeaway was that the defenders didn't have the same capabilities as the offense.
Okay, now you would look at this and say the airbags were off because this was an experiment where the agents, you know, the guardrails were down, the airbags were off, and ultimately they found the way out of the sandbox. But the reality was Hugging Face, you know, when the 1st thing they did was they went to the frontier models and they said, what, like, is this thing real? What's going on? And they got refused, right, because the airbags weren't over there.
So what did they have to do? They actually had to go to open weight models to kinda sift through all this data and figure it out. So a big piece for me is t is to basically make sure that defenders have the same level of a frontier defensive AI as as what the offense is, uh, is using. And that was a big takeaway, uh, for me, and and part of the things that that CrowdStrike is working on.
And is that part of the context that you mentioned of, um, the guardrails being so strong on the state-of-the-art labs models that to some extent, you could argue there aren't that s those same defensive capabilities available to every enterprise? In
That's correct.
other words, the fact that they had to go to open weight models kind of suggests that you can't actually do it currently with the labs. And so the question then how do you adopt that? Yeah,
Yeah, yeah, I mean, like, you, you try, you try to make a PowerPoint in in with a frontier model, uh, around cybersecurity, and it refuses you. So, you know, this is part of why we created SafeMind and and our and our red and blue models just for that reason, because we wanted to give, uh, frontier caliber defenses, uh, to our customers that allow them to operate without sort of all these sort of refusals. Um, and and I think that's gonna be an important element going forward for for many companies.
I think, um, I guess back to the offense versus defense side, it seems like with every technology shift, offensive capabilities are are always a little bit ahead. When you have a big change in how things are done for a few reasons, I mean, one of them is just with offense, it can be very directive. I need to go do X, while with defense, it's a broad surface area you need to cover, so to some extent, you need to cover more stuff.
And then secondly, it's just technology adoption and diffusion, right? You are one atta- you are a set of attackers, you know what you're going after versus the diffuse ecosystem of enterprises that you need to defend against. How does CrowdStrike think about ensuring that those defensive capabilities are very broadly available?
Well, this is one of the things that we focused on, uh, we created our own cyber super intelligence labs before the term was was popular, I might say, um, so I guess we we kinda nailed that one. But essentially what we've done is we've created our own basically frontier s type type lab that is totally focused on, uh, security models, both red and blue and the harness, and we call it SafeMind as a system, but we've got Red, Tempest and Blue Solano, those are the two models.
Now, we worked extensively with NVIDIA on this, and, uh, we basically levered we leveraged a few model types, but NemoTron is one of the main models that we leveraged on on the blue side. And when we sat down with them, and I sat down with Jensen, we've been working on it for the last six months, and we kinda went through how this would work. And essentially, to your point, like, you know, the defense always seems to be one step behind.
So the big part of of what we're doing is we're iterating the red and the blue. So for our customers, we have a view of what their environment looks like. And then a and this was Jensen sort of i insisting, like, you have to have a digital twin. If you look at other industries, they have digital twins. And he goes, you have to have one in security. So we built a digital twin where we can take all of the information that we know about a customer's environment.
We can basically recreate that in a virtual world, and then we can iterate the models and loop the models. Where red continually is focused on getting in until it runs out of turns because blue Solano has actually figured out what red was doing. And that's really the way I think we have to keep up with, uh, the threats, because red is constantly iterating and you're gonna have to have this loop in place.
Uh, that makes sense. I I mean, even just from a classic cybersecurity perspective, like knowing your environment is one of the, um, only advantages that you have as a defender. As well as as as CrowdStrike has pioneered, it's like, you know, the ability to share intelligence and understanding of the adversary, right? Um, actually, I I think it'd be useful to just like zoom out and describe, like, w- uh, you know, the the uh, key offerings and investments on the R D side that CrowdStrike is making today.
Because I I think like it's a it's a massive cyber player, um, but I think the the root of like endpoint protection and understanding the adversary might be an outdated view.
Yeah, I think that's right. You know, we, we've really, you know, probably over the last, you know, four or five years, I guess, really moved into, you know, much broader platform capabilities. We've got 33 modules today, and, you know, ultimately, when I started the company, we had one module, and now we've got 33. We've built capabilities not only across endpoint, but around cloud, uh, IT identity, um, and many different areas that are out there. So if you look at where we've made a lot of investments in some of the things that we just talked about at our latest user conference, FalCon, um, it was around what we call AI DR.
So we pioneered EDR endpoint detection response, which a lot of it in the early days was a computer and a human. Um, and then, you know, we've added obviously many more things around that and prevention and all those sort of things. But when you look at what an agent is doing, it's it's essentially a superhuman. It has an identity, it has access to compute, it runs on your computer, uh, or it can run in a container.
Um, it has access to data, it has networking capabilities, right? So what we wanted to be able to do is to 1st find all of this rogue and shadow AI, cause nobody knew where it was. Like, back to your point, you gotta know what your environment looks like. And then 2, we can instrument from the birth of an agent, what that agent does. So if you're in, you know, Claude, Claude Code, you start an agent, it has sub agents, it has tool calls, it, you know, goes to an MCP server, it pulls data, whatever it's doing.
We can actually tell you the entire chain of events, and that's incredibly important and then be able to add policy and blocking rules around those. Um, so we've spent a lot of time in that area, and I think from a business perspective, if you look at the stats right now, on average today, which I actually think is low, but they say you can have 90 different agents, you know, as a as an employee. Um, well, that can be a much bigger opportunity than just EDR for us.
So that's one area that we've been focused on. The, um, the 2nd area then, as I mentioned, is creating this super intelligence lab and creating our own frontier models. Again, our models, we, we, we start with existing models like, uh, a NemoTron class, you know, mostly the 3 class of NemoTron. We're helping them sort of fine tune some of the things we need for the force that's coming out. Um, but then, you know, we kind of, you know, post train them in and using all the valuable data that we have over the last 15 years to get the best outcome.
And I, uh, and I'll say one more thing, and I'll pause. The the harness makes a huge difference. Like almost any model we take, whether it's frontier, whether it's just an off-the-shelf, uh, open-weight model, we can get much better results at a way lower cost because of the m the harness that we we created that goes with it. So those are some of the big investments. Um, you know, there's more around identity. I'm happy to cover any of those, but let me pause there and and see if you had any any other comments or thoughts on those.
That was gonna be one of my, um, big questions about your lab's investment here, which is a I I think a reasonable question to ask about any model training effort is like, why isn't that just a frontier? Um, y- you know, you work with OpenAI and Anthropic, uh, like, capability, um, uh, these these labs, of course, uh, they, um, realize the importance of security. And then they have these programs for like vetted defenders, be it like Daybreak or sort of Mythos access.
Why is that insufficient? Uh, I mean, you have a you have a capability argument based on, you know, harness your ca- your understanding internally. Data is there. Is there anything else structurally that you think is important for an independent security vendor here?
yeah, I think there's a few things there. So 1st, what what the labs are creating are is incredible. I mean, where they've come from, what, what they created, it's unbelievable, right? And we also wanna make that available to our customers. So the general theory is if you have SafeMind and you wanna leverage a frontier model, great, we can, we can leverage that. Um, if you wanna leverage our models, fantastic, our models will be significantly cheaper. So just to give you an example, um, so f- for a general remediation, if you use like, uh, uh, uh, the highest caliber frontier cyber model, probably cost you about ten.
Ours is like $0.03. So there is a huge cost difference, um, which is important. But then when you look at, you know, our model, be because of the way we've post trained it on our data, a huge difference is we have the data. Like, there is no Reddit of CrowdStrike security data. You can create synthetic data, but it's it's really not the same as what we have. We got 15 years, uh, well, maybe a little bit less because I started 15 years ago, but, you know, call it over a decade of security data, including from our MDR, our managed detection response service that has all of the sort of threat and and vulnerability pairings and sort of the attack chains.
And we've basically created this massive amount of data that we use to train it. So we g- we're getting great results on it, and, um, I think it is very focused be because of the post training. You know, again, it's not gonna cure cancer, but it's gonna give you great results for security. That's number 1. Number 2 is there is a safety and trust aspect. We have all of the data from our customers. We've been entrusted with that for the last 15 years.
We know how to protect it, we know what they want, we know how they want to use it. W- they there there's a real concern by customers. They don't want all this data going out to the labs. They don't want all their sensitive information. They don't want it their information being trained on, right? So we can provide a a way to get them the trust and security that they want, um, for for their own security program.
And this this sort of concept of AI sovereignty is very important to many of our customers. So, again, we're not saying exclude the labs because they have incredible technology, but you can pick what you want or you can route between, you know, the various models. But ultimately, we wanna give customers choice.
I think you've also talked a little bit about separation of church and state, you know, the you wanna make sure that to some extent, there's a neutral third party that can also, uh, you know, interrogate some of the models that the labs are providing or some of the other open source out there. And, you know, it seems like CrowdStrike can definitely play that role as well. So, um,
I think the number one factor in security is is is being trusted, being a trusted partner. You know, you can look at the vendors out there, there's, there's obviously a few big guys that are that are out there, there's many security companies, but, uh, trust is is, uh, built in drips and lost in buckets. And, you know, cus- customers, they they don't want sort of cowboy activities that may put their their data at risk. And, um, y- you know, again, I and everybody's do trying to do the best they can.
But I our whole focus, right? And I think focus is incredibly important in any business. Our whole focus is security, and our whole focus is keeping our customers safe. And we're not, we're not trying to solve the world's problems in in AI and other areas. So I think that is is certainly an advantage for us.
I if you were to think out two years from now, what do you think is the state of the security world? Like, what do you think has shifted from this point forward?
Certainly, within two years, you're gonna have local models that are all running on people's computers. Um, you know, you may have a mixture of local models, plus sort of last mile routing into into frontiers in terms of what they look like. You're gonna have a lot more open weight models. Um, you know, you can just see the shift in the tokens, you know, over the last six months.
What what is the shift that you're seeing, and how what's the order of magnitude there, and how long do you think that continues for?
I talk to CIOs, you know, almost every day. I had dinner with one tonight at a f- big bank in New York. And and basically, the comment is, you know, we, we, we got into AI, we went all in, great. The costs were out of control, you know, token maxing, you know, stopped when people got big bills. And, you know, I said, Well, what, how do you think about this? Like, what are you gonna do? And they said, the models are so the open weight models right now are so good, even if they're three months behind, it's plenty for what we need.
And one of the customers comments, and and they know, uh, I'm into Formula 1. They said, any Formula 1 activity that we need for the bank, we'll go to the frontier. Like, if it's alpha and we're making money and it's like, you know, super important, we'll pay the money. Not no big deal. But for all the other things, like, it's really expensive and we're gonna create all of our own open weight models, you know, and do more post training on them and those sort of things.
That that's the way they're gonna operate, and I firmly believe that. I'm seeing that in customer after customer, that that's, that's, that's the way they're gonna go. And I think, you know, part of some of the challenges in the current environment is I think people are are all seeing that. And, you know, there there is some concern, you know, at the frontier level of of what's gonna take place.
And so, I guess on the open weight model, one of the, uh, statements that tends to be made is that they are at least in part distilled from the labs. In other words, you have Chinese open source players who are effectively training their models against that Anthropic or OpenAI model or, you know, other sort of frontier models. Uh, and many of those labs, you know, are trying to roll out anti-distillation efforts. So they're trying to build into the model the capability to not get distilled.
Do you think it's just sort of a back and forth where the capabilities keep going, almost like offense and defense in cyber, or what's your prediction in terms of the state of distillation in two years, or, you know, within some reasonable timeframe?
I mean, I guess my 1st comment is what were the frontier models distilling when they got all that information? I mean, we know all of the, you know, sort of the lawsuits around where the data came from and and how that unfolded, right? So we kinda start from that.
So it's like a fine liquor, it's triple distilled, you know, so it's, uh, It's triple distilled, exactly yeah,
like, so where, where, where does the distillation start from? You know, um, so I think it it is what it is, but the the the labs certainly are focused on on trying to prevent distillation. And and obviously, you know, when you look at the Chinese models that are out there, there's a lot of distillation takes. There's factories of systems that are, you know, bouncing all over the internet, distilling this. I I think the challenge is the the genie's out of the bottle.
I mean, there there's a lot of distillation already taken place. The models are very good, but we also have to give credit where it's due. If if you don't have all the chips and you have to be resourceful and you still wanna have top models, you know, I think there's some innovative things that are taking place outside of the US. and we gotta be mindful of that, right? Um, so yes, it's hard. I I think it's gonna be hard to put the genie back in the bottle on distillation.
They they're gonna do the best they can, but, um, you know, a lot of the models are are, you know, I I think now, you know, what does it look like in two years? I don't know. You you've got great models today, they they can last for a while, and you have a lot of data that's already packed into them. So we'll we'll see how it unfolds. I I wish I had the crystal ball on that one.
oh, no, I was just curious. And then I guess another, it's always hard to predict these things. But I'm a little bit curious, do you think there's structural advantages for Chinese companies to stay at the frontier of at least open source and open weights, or do you think, uh, US. open weight models are likely to catch up in the near term?
Well, there are some open weight models coming out. I'm, I'm sure you, you know, you may be invested in them as well. Um, I I think having great US. open weight models is gonna be very important. So, you know, one of the reasons why we worked with with NVIDIA was, of course, NemoTron is open weight and open source, not just open weight, but having other alternatives is gonna be very important. And I think there's, you know, again, just being in technology long enough, you kinda look at how things unfold.
Like, the the operating system of the internet is Linux, like, period, okay, and there's a reason for that, and there's gonna be an operating system of AI, and it's pro. In my mind, it's gonna be, you know, an open source, open weight. When that happens and how it all unfolds, I mean, we'll see. And it doesn't mean there isn't room for any frontier model. It doesn't mean that people are not gonna make a bunch of money off it.
You know, the application stack sort of around it, um, you know, the harnesses and how things are connected, and how you think about enterprise, all those sort of things are gonna be really important. But ultimately, like, there's just too much cost involved right now. And and when you look at a lot of the these frontier models, I mean, they're being commoditized. If the Chinese models are three months behind, you know, a lot of companies are going like, three months is not what, what is, what am I gonna get out of it?
Th- I does everything I want, and that's the challenge.
A a top five financial said to me, um, uh, recently, like a, you know, a large technology buyer there, um, that, uh. There was no chance that their organization had the capability to absorb a new model release every ten days anyway. And Correct. so three months behind, you know, was, uh, uh, faster than the acceptable, like, regulated pace of their organization, so it made no difference to them to to to some degree. Um, I I think open is here to stay is a big part of the puzzle.
Uh, you know, one one question. I I know we're gonna keep asking, you know, look into the future, and it's kind of impossible. But given your history in in security, this is the 1st time in my career in technology where some very, uh, smart, well educated technologists are like, we need to turn the internet off. We just gotta air gap everything because the internet is gonna fall apart in the next year and it'll be unsafe for critical systems to touch it.
Um, how? How much do you believe that to be like a reasonable probability in the future, where there are lots of critical systems that society cannot rely on touching the internet?
I mean, I think it's nonsense, you know, uh, it like, is security and and what agents can do and and the risk that they pose is that serious? Of course, that's serious business. We should all look at that and go, Yeah, we gotta make sure that we lock things down, you know? But is the world gonna end because of these agents anytime soon? No. And when you look at what the agents are doing, I I think there's this sort of mythical view of what an agent does.
Like, an agent doesn't just beam itself from your computer to another computer. It actually just does the same thing a human would do, just better and faster, right? So if you think about what companies like CrowdStrike have done for a long time, we prevent against these sort of attacks. Now they might come faster, they might be, you know, more lower level sort of vulnerabilities chained together and those sort of things. Great, um, you know, great for the adversary, not great for the defender, but it's building the technology that, um, can deal with all these sort of attacks.
So what has AI done? It's created more vulnerabilities, so you have more of them, and, you know, you have more exploitation coming faster, so it's more and faster. But it there isn't, like, a unique class that all of a sudden an agent just beams itself somewhere. And I think people just get wrapped around the axle of, like, there's some new class of technique that's taking there, there isn't. It's just been trained on what we know, it hasn't developed its own yet.
And that's why I kinda look at it and go. Security again is gonna be an enabler for AI, which I think is incredibly important for the security industry. Cause normally it's, you know, you're putting the brakes on, you know, we have to take time. You're not letting the business run as fast as they want. And now everyone's coming to to CrowdStrike and and others going well, how do you make AI secure so we can go faster in AI?
So that that's the way I would look at it, like it AI is here to stay. I think some of the the the concern is there, and it should be there, but sort of like, there's a lot of, um, uh, kind of doomerism around some of this stuff. And I, I just don't think it's it's necessarily warranted.
I know you're also, um, you know, sometimes involved with companies as an angel investor and, you know, other parts of the ecosystem. Are there other areas that you're very excited about right now in terms of key areas of change or focus in the AI world, in addition to the security world? Like, what are some of the things that you think are really exciting that either just happened or are coming?
I think, you know, chips used to be kind of not in fashion. Obviously, we we know they are in fashion today, and you're only gonna need more of them and faster of them, and and more specialized. And I I do think there will be some big breakthroughs in the chip world. You know, when we think about just sort of the speed, just take, you know, Moore's law might be a bit data but dated. But at the end of the day, you know, they, they, the chips came out and things just got faster and cheaper.
I don't know the trajectory, but I can guarantee you it's gonna be faster and cheaper in the future like that, I I can see. So as these specialized chips are developed that, you know, focus in maybe s particular areas, as new science is developed around the mathematics. Um, you know, there can be step function increases in performance, you know, uh, less power, those sort of things. So I think chips is a big one, the the open weight models.
Um, I think there's a lot of room to grow, certainly the US. models. And I do think there's gonna be a lot of advancements in taking these models and shrinking them down, you know, and and and being able to run them on your phone as an example. So I I mean, we probably all know the history. Remember the days where, you know, it took, uh, a football field-sized room of computers to to, you know, launch the space, you know, 1st rocket.
Like, we have more power in our iPhone today than than they had in the entire room. So that's gonna come. And I think the area where you can make AI faster, cheaper, more efficient. And and you bring it to the desktop and you bring it to the the the phone where a you're, you're doing it at the edge. I think all of those are sort of exciting opportunities because that's really where I see, you know, you're, you're probably gonna have a hybrid model, but you can do a lot of it now locally.
And again, just looking at the, um, you know, the TPUs, NPUs, et cetera, like these are, they're, they're all on systems that we have today, and and a lot of times are going to waste. But I I don't think they're gonna be going to waste that that much longer.
Just because it is, you know, your job to think ambitiously about what the defenders can do all day. Uh, I, I wouldn't say like, AI, def- uh, s sorry, Defender, security defenders are, uh, doomers in that specific term, but they're like, generally kind of dour and like, pretty pessimistic about the, um, cat and mouse game getting any better in their favor. Can you can you make your best case for like, actually, you know, if people invest correctly and we make better products and we automate many cert- like, is there a version where it is actually defense dominant over the next decade?
I think it can be because if we again, just with with the concepts that, uh, you know, are, n- they're not necessarily new or unique. But the the the ability to always rapidly understand what the offense is doing and making sure that your defense is keeps up with it is is gonna be the key. I mean, it, it, it's sort of the the modern day version of the the hacking exposed book that that I co-authored many years ago.
Which was like, if you show people how the offense works, the defense is gonna be a lot better. Now, that was a much slower sort of tempo than it is today, but I think that's gonna be important. And as an industry, making sure this is a key thing for us. And and I really drill this home with with our our team at CrowdStrike is you've got to give as good or better capabilities to the defender than what the offense has.
And if you're just slightly better, you know, just mathematically, you're still gonna have an advantage. And I think it's again, it's always gonna be a cat and mouse game, but being really smart about it is gonna be important. The only way to fight AI today is with better AI. And, you know, you, you've, you've gotta u- use the advantages you have. The defender knows their own systems, the defender has the architectural diagram, the defender has the code, right?
So how do you put all of that into something that you can dramatically increase your probability of outsmarting the offensive AI?
The only tactical thing I I think is an interesting idea is the idea that, um, you could over time, if if it is going to be mostly agent written code, you could enforce better security architecture. Um, and, you know, perhaps fewer vulnerabilities. Or a lot of people where, you know, formal verification has always been enormously niche, right? But even even short of that, like, okay, why is all the software insecure? Because, you know, one one big reason is that engineers are, um, you know, there's a limited number of them with limited capacity, and we're all trying to move very quickly.
But if you can scale at the speed of compute or your willingness to spend, and agents have infinite patience, like, maybe we could actually end up with more secure software over time. It's interesting.
I I think there there is, like, the industry's at a point now where where the code is is probably good enough where you're gonna get definitely some benefits. But again, the code was was also trained in other area, like it's was trained on stuff that, you know, may have had issues. So, uh, you know, I, I, I have my own developer license, I, I build my own sort of apps and and things like Claude. And and, you know, I have my own security agent that reviews the code.
And it's like, well, Claude just created the code, but then Claude is now with its security app going well, that code is insecure. Well, you just created it, like, why just do that in in the 1st part? You know, maybe there's some optimization, but I think, um, what people don't necessarily realize, uh, is that even with with zero vulnerabilities, which I hope the industry can get there, you're still gonna have breaches. Cause a lot of what we see people today doing is they're not necessarily breaking in, they're logging in.
There's such an issue around identity. But at the end of the day, if we can get better code and we can have less vulnerabilities and we can ripple that through, that'll be great. The the challenge is it it's such a long tail. You take Microsoft as an example, just because of of all of the the um, you know, compatibility that they want to maintain going backwards, backward compatibility. You know, there's things that never kind of get fixed or or modernized.
And, you know, that's just one important element in our ecosystem. But think about all the other software that's out there. You're not gonna just fix it all overnight. So it's gonna be a really long tail.
You've been running CrowdStrike for, uh, 15 years, and before that, you were an entrepreneur. Um, uh, security, I I've always loved security because, like, we're never as a business person, it's both an interesting and important mission. As a business person, like, this problem is not going away, right? Um, but, uh, I I feel many com- many CEOs of large established public companies, uh, are feeling a little bit unmoored in this period of time. Because so much of the value that they have built, uh, in their incumbency is is being challenged.
Like, how do you, how do you, you know, think about keeping yourself up to date, keeping the organization agile, continuing to win? Any any general advice on that?
yeah, hap- having to jump into that. So 1st, I think you have to recognize it's a different operating model and cadence. You know, w- we've got thousands and thousands of people, right? And when you think about some of the AI native companies that just started, they have a handful of people. You know, Claude is running all day. They've got agents doing all kinds of stuff. Like, that's the that's the new world, which is great. But you, you can't be in a, uh, not that we are, but you can't be in, like, the waterfall mentality of development.
Like things pop out a year later, like you're gonna be out of business. So I think you 1st have to realize that having these massive teams and these long schedules like that is not gonna cut it. So having smaller teams that are, you know, squads or, you know, smaller groups that are focused on a problem, they can leverage AI. But you can't have, like, thousands of people show up to solve a problem, you gotta have a smaller team to be able to do that.
And we've done that, we've done that with, like, our Falcon Guardian product. And to your point, like, I was personally involved, you know, had three day a week meetings, I was involved in the design of it. It's like, here's what we're doing, like, change this, change that. And you gotta get back to your like startup roots because it it it takes that level of change, you know, in in a bigger company. And there are many companies that are much bigger than we are.
And, you know, you just can't be slow and plodding. So that's number 1 is to realize that the operating model and cadence is changing, and you gotta change what you gotta disrupt yourself internally. And then you just have to look at, like, what, well, what are your moats? Like, if you're a system of record, it's a pretty big moat. We're we're a security system of record, right? And we have everyone's data for many years. And by the way, we're net data creators, our our sensors, we call them security agents, right?
But I mean, like, not to be confused with AI agents. But the software that that endpoint agents, right? That is a net data creator. So we're creating 7 trillion or so events like, uh, a day. And if you're constantly creating this new data stream, that is gonna give you an advantage. Because what's the lifeblood of any AI system is the data that you feed into it. So I think, you know, for us, it's always like, well, what are our moats?
How do we look at it? How do we make sure that, you know, we we disrupted the incumbents, we disrupted the semantics, the MacAfee's, you know, all those other players are, like, all gone for the most part, or changed or bought or whatever. And you never want to be in a position where somebody, you know, disrupts you. So we're always like, how do we disrupt ourselves? how do we do things differently? how do we think differently?
But the end of the day, you know, if my my main focus, and I tell this to our team all the time, if we take care of the customer, the rest takes care of itself. What the customer wants is they want to be protected, they want to have trust in their in their vendor and their partner. And at the end of the day, they want to make it super economical. Like no one wants to spend unnecessary money.
So give them the best outcome at the best cost and an experience that doesn't take, you know, seven consoles and six chairs to run around in and just make it easy. And that's the number 1 thing, you know, that our customers always tell me is like, it is so easy to deploy your stuff. Just we're at a meeting this morning and it was like, you know what, we're this two companies merging together, they're gonna go with us.
Um, the other company didn't have us, and it was like, your stuff just is so easy to deploy. And I think that's a big one. Is time to value?
So a lot of people right now are talking about job losses due to AI, although that really hasn't happened. I think the opposite has happened in terms of net job creation so far. And similarly, there's a lot of questions about should my kids study and what will happen in the future for jobs and the like. What what is your view on that, or where would you recommend people go going forward? Or, you know, is this even a long-term issue?
Well, just in general, from an AI perspective, and if you look at the t the technology changes, um, that have taken place over the last couple hundred years, like each major shift has created more opportunity and more jobs. They're different, but there were more of them, and I firmly believe that's gonna be the case here. And, you know, when you look at whether it's programming, you're still gonna need people to manage the agents. You're still gonna need people to think about these architectures.
It's not just fire and forget. I know a lot of people would like to say that, but there is human creativity that you're gonna need. Uh, I think specific to cybersecurity, this is one of the areas that that I always talk to to young folks about, people in in, uh, in university. Is it's a great field to be in if you understand AI, if you understand security, it it's an ever changing field, it and it's it's actually not a solvable problem.
I think every vendor would like to say they solve the pro- it's not solvable. And you, you've gotta mitigate the risk, you've gotta come up with the best solutions, you've gotta be able to protect the customers. And, you know, there's a lot of things that go into it, but it's always evolving. There's gonna be technology in five years that we haven't even thought about today. And if you're in university, you're gonna be some of the people that are gonna be thinking about how you protect it.
So I always, you know, try to give back and think about the d what college kids are doing. Or I get asked all the time from parents, what should what should my kids study? I think cybersecurity and technology and AI is a fantastic field.
It's a great note to end on. Thanks so much for doing this, George.
Thanks for joining us.
Oh, you're quite welcome. Thank you so much.
Find us on Twitter at No Priors Pod. Subscribe to our YouTube channel. If you wanna see our faces, follow the show on Apple Podcasts, Spotify, or wherever you listen. That way, you get a new episode every week. And sign up for emails or find transcripts for every episode at no dash priors dot com.