新一代顶级掠食者：CrowdStrike CEO George Kurtz 谈手握 AI 智能体的对手

For the first time, safety is the gas pedal, not the brake.

How would you describe what you see AI doing for adversaries?

1. Agentic Landscape

The new apex predator is the agent state. The same level of capability that a nation-state would have had just a few years ago is essentially what a hacktivist has today. This has dramatically changed the overall threat environment.

Everyone comes to CrowdStrike and others asking, “How do we make AI safe so we can move faster in AI?”

The only way to fight AI today is with better AI. The defender knows his own systems. The defender has an architecture. The defender has code.

How do you translate all of this into something that can significantly increase your likelihood of outsmarting the AI of an attacking actor?

George, thank you very much for doing this.

I’m glad to be here.

2. AI and Adversaries

This is a time of dramatic change in technology, and agents are taking over the business. CrowdStrike is at the forefront of this. What can you tell us about what you’re seeing in real life right now? I think you said there are 2 to 3 times more agents taking action on people’s behalf than they realize.

3. Adversary Geography

What we’ve seen in the business has been incredible. If you follow the trajectory of the frontier labs and look at their revenue, you can see how many more agents are emerging, whether it’s Claude Code or Codex. They’re all over the environment. Obviously, there are many other types of agents, but those are useful benchmarks.

We’ve seen them popping up everywhere, and part of the problem with these businesses is that they have no idea where they’re popping up. Shadow AI has become a huge problem for many companies. I know we’re going to talk about what agents do and all the bad things that happen, but overall, understanding where their AI is going, who’s doing what, and, more importantly, what identity they’re actually operating under has been a big challenge in the industry and a big focus for us at CrowdStrike as we try to crack the code for our customers.

How would you describe what you see AI doing for adversaries? You’ve described this as flattening the pyramid of adversaries.

When I think about adversaries, I call it the pyramid of pain. At the bottom of the pyramid are the hacktivists: the lowest level of difficulty, but the largest number of adversaries. In the middle of the pyramid is cybercrime: more complex, and a little smaller in number than the hacktivists. At the top of the pyramid are the nation-state actors: the most complex, but relatively the smallest number.

That’s been completely blown out of proportion, because what we’re seeing now is that the new apex predator is the agent state—not the nation-state, but the agent state. The agent state has taken all these very complex techniques, different methods, and costs and flattened the entire pyramid. The same level of capability that a nation-state would have had a few years ago is essentially what a hacktivist has today. This has dramatically changed the overall threat environment.

We’re seeing more vulnerabilities discovered and faster attacks. When attacks happen, they happen one after another. You can tell pretty clearly now whether it’s an AI attack because you can see how quickly it repeats its actions. At some point, though, they’ll introduce a low-and-slow approach, and it will be harder to tell what’s AI and what isn’t.

This is what we’re seeing, and I think it’s one of the real challenges organizations face: a barrage of attacks that have dramatically increased in sophistication. Ultimately, if you give an agent a task and enough time, and the agent either needs to get inside or it’s going to die, it’ll try to find a way in, and often it’ll be successful.

Do you see any changes in how these attacks are carried out by the enemy? I know we talked a little about open-source models versus closed-source or open-weight models. Do you see a change where most of them are open-weight models? Is it a mix, with people using state-of-the-art labs and capabilities? What are people actually using to carry out these attacks at the moment?

We see a lot of this on open-weight models. If you look at the latest generation of safety guardrails on any of the leading models, to be honest, you’re unlikely to be able to create a prompt about cybersecurity. It literally rejects you and downgrades you to a lower model.

What we’ve seen is that if you look at GLM, Qwen, or any of the existing models, they’ve been shown to be very effective in dealing with cybersecurity. Many models are either abliterated or specifically created for cybersecurity. These can be used for red teams, and they’re the type of models you can just run.

I have 2 DGX Sparks at home, and I have all the models running. Anything I want to do, I can do. Want to create a ransomware program? No problem. Want to create a ransomware kit? No problem. Want to deploy? Want to iterate? It’s all there for you, and it’s one step away from running all of this on your laptop.

Think about where the Mac is today, even the new Macs that are coming out. Mac Studios will be incredible AI machines. The tip of the spear is that you can run without restrictions. The frontier models are somewhat useful if you need PowerShell scripts or something like that, but of course you can do a lot of that with open-source models right now.

Is there a common geography? Do most attacks come from one region? Is there a common type of attack? Is it mostly ransomware, or something else? Are these phishing campaigns? I’m curious whether the composition has changed as you’ve seen the introduction of AI tools.

We’re looking at it from the adversary’s perspective, so let me reframe the question a little bit. Is it a nation-state? Is this cybercrime?

4. Attacker Capability in Context

With nation-states, we came up with these names—bears and pandas, depending on whether they’re from Russia or China; they could be from Iran, kittens and so on. We’ve seen a lot of attacks. We’ve seen how busy North Korea is with these attacks. They have to bring money into their economy, so they’ve been very busy stealing information. A lot of this is related to ransomware and getting money.

We’ve also seen a lot of activity from China using artificial intelligence, and we’ve seen it in the cybercrime world. When someone gets into a system, we track the time of the breach: if they get into your system, how long does it take them to break out, pivot, and move to another system? Usually, the first system they encounter isn’t the most interesting. It may be, but it’s not the most interesting. The crown jewels may be elsewhere.

On average, it used to be days, then hours, and now it’s minutes. On average, it’s 29 minutes, but we’ve seen it as short as 27 seconds. In fact, we’re now observing the speed of execution. As soon as they connect to the machine, you can see them iterate very quickly.

People should remember that these are no longer malware, but TTPs. They can essentially create TTPs that produce a completely unique signature of the direction they’re going, how they’re gathering information, and where they’re going to pivot. That makes it harder to understand, but it’s exactly what we’re seeing now. It’s true for everyone, from nation-states to cybercriminals.

You’ve been a leader in security since Foundstone and the founding of CrowdStrike. Can you put the current period of growth in offensive capabilities, and perhaps defensive capabilities, into context? Do we now live in a world dominated by the adversary? How do you view this problem in general?

I think this is one of the most exciting times in technology. I’ve been doing this for over 30 years, and I always like to say that security follows the slope of the technology curve. When I first started in security, it was a Cisco router and maybe a website, right? There really were no firewalls. There were these toolkits, the TIS Toolkit and other things. I may have made some myself, but there wasn’t much there.

Now, think about the complexity of where we are today, with the cloud and all the other technologies related to artificial intelligence…

I’m not going to go into all of them, but it’s more complex today than it was back then. So you need all these small companies, including the big ones, that are able to protect all these technologies. When we think about where we are today, that’s actually the most interesting thing, but it also creates the biggest opportunities, because opponents usually find gaps or seams. They find new technologies that may not have good technologies to protect against.

5. Hugging Face Takeaways

They use things like npm packages, right? In the old world, as you well know, no one in corporate America just downloaded something and ran it on their computer. It had to be checked, scanned, and all that. And now we release an agent, give it an identity that is essentially our own or worse, and then we release it and say, “Do whatever you want.” And by the way, download any package you want to download. Obviously, these are the things that generate a huge amount of exposure.

6. George Kurtz Introduction

So we kind of handle any production code as soon as it comes out? No. They can’t cope. You see some of these things after a while and you scratch your head. You shouldn’t be surprised. It was something like, “Let’s give him a task and we’ll come back in a few days and hope it worked.” Nobody is following this. Nobody understands what it does, and that’s problematic.

7. Open Weight Models and Adversarial Attacks

We can talk about some kind of explanatory power for what these agents do, which would be really important. But now it’s like a black box and no one knows what’s going on. They’re hoping that something will happen and they’ll show up on the other side. But now is a great time for technology. Artificial intelligence will fundamentally change the world, and cutting-edge labs are doing amazing things. But ultimately, it all comes down to the security that such technologies provide. For the first time, safety is the gas pedal, not the brake. And this is what we see.

When did you realize that this tipping point was happening, at least in the security world? For a lot of people, the OpenAI issue with Hugging Face was a big moment where people realized there was a big shift happening in terms of capabilities and technology, and what could happen. Your point is about monitoring and so on. Was that a defining moment for you, or what were the defining moments?

Let me ask you a question first. What’s your main takeaway from Hugging Face?

That anyone running advanced models must be careful about the quality of their environment and have very good ongoing monitoring of production.

Yes, and this is the main conclusion. But I think many people would say that this was the first agent-based end-to-end attack. All of this is true. You look at what these things did, and people talk about them hiding and trying to survive until the next generation, and so on. All of this is true. My most important conclusion was that the defenders did not have the same opportunities as the offense.

8. CrowdStrike Frontier Security Models

You would look at it and say the guardrails were off, because this was an experiment where the agents’ guardrails were down, and eventually they found a way out of the sandbox. But the reality was harsh. The first thing they did was turn to the models at the frontier and ask, “Is this real? What’s going on?” And they were refused, because the guardrails were there. So what did they need to do? They had to turn to open-weight models to sift through all this data and figure it out.

It’s important to me to make sure that the defenders have the same level of AI on the frontier as the offense. This was an important takeaway for me, and it’s part of what CrowdStrike is working on.

Is it difficult, given that you mentioned the guardrails are so strong on the most modern lab models, that to some extent it can be argued that the same protective capabilities are not available to every enterprise? In other words, the fact that they had to move to open-weight models hints that you can’t really do that with labs right now. So the question is, how do you use it there?

Yes. How do you try to make a PowerPoint presentation with a leading model about cybersecurity and have it refuse you? That’s part of the reason why we created SafeMind and our red and blue models: we wanted to provide our customers with advanced protections that would allow them to operate without all of these failures. I think this will be an important element in the future for many companies.

Going back to the offense-versus-defense confrontation, it seems like with every technological shift, offensive capabilities always get a little ahead when there are big changes in how things are done, for a couple of reasons. One is that on offense, it can be very directive: “I need to do X.” In defense, it’s a wide area that needs to be covered. To some extent, you need to cover more things. And secondly, it’s just the introduction and dissemination of technology, right? You have one attacker, or a set of attackers, that you’re chasing, against a diffuse ecosystem of enterprises that you need to defend. How is CrowdStrike thinking about ensuring that these defensive capabilities are very widely available?

That’s one of the things we focused on. We created our own cyber superintelligence labs before the term became popular, I would say. I think we’ve kind of managed that, but essentially what we’ve done is create our own lab, a frontier-type lab that’s completely focused on security models: red, blue, and harness. We call it the SafeMind system, and we have Red Tempest and Blue Solano. These are 2 models. We’ve been working a lot with NVIDIA on this, and we basically use a few types of models, but Nemotron is one of the main models that we use on the blue side.

9. CrowdStrike Key Offerings

When I met with Jensen, we had been working on this for the last 6 months, and we roughly discussed how it would work. To your point, the defense always seems to be one step behind, so a lot of what we do is red and blue iteration. For our clients, we have an idea of what their environment looks like. Jensen insisted, “You have to have a digital twin. If you look at other industries, they have digital twins. You have to have one for protection.” So we created a digital twin where we can take all the information we know about the client’s environment and recreate it in the virtual world.

Then we can iterate the models and loop them, where Red keeps focusing on the input until the queues run out, because Blue Solano has actually figured out what Red was doing. That’s how I think we have to keep up with the threats, because Red is constantly iterating and you have to have that cycle in place.

That makes sense. Even from a classic cybersecurity perspective, knowing your environment is one of the few advantages you have as a defender. And, as you know, CrowdStrike pioneered the ability to share intelligence and adversary understanding. I think it would be helpful to step back a little bit and describe the key offerings and the R&D investments that CrowdStrike is making today. I think it’s a huge cyber player, but the root of endpoint protection and understanding the attacker may be an outdated view—not just endpoints, but cloud, IT, identity, and many other areas.

If you look at where we’ve made a lot of investments and some of the things we just talked about at our last Fal.Con user conference, it was around what we call AIDR. We started with EDR, endpoint detection and response, which in the beginning was very much a computer-and-human effort. Then we obviously added a lot of other things related to that, prevention and all that.

But when you look at what an agent does, it’s essentially superhuman. It has an ID, has access to computation, runs on your computer or can run in a container, and has access to data and networking capabilities. We wanted to find all this rogue and shadow AI first, because no one knew where it was. Going back to your point, you should have known what your environment looked like. Secondly, we can instrument what the agent does from the very beginning of its creation. If you’re working in Claude Code, you run an agent, it has subagents, it has tool calls, it contacts the MCP server, it gets data. Whatever it does, we can tell you the whole chain of events, and that’s incredibly important. Then we can add policies and blocking rules around them.

10. What Differentiates CrowdStrike’s Models

We’ve spent a lot of time in this area, and I think from a business perspective, if you look at the statistics now, they say an employee can have 90 different agents on average today, which I don’t think is enough. This could be a much bigger opportunity than just EDR for us. That’s one of the areas we’ve focused on.

The second area, as I mentioned, is building a superintelligence lab and building our own models at the edge. Again, we start our models with existing models, like the Nemotron class—mostly the 3 Nemotron classes. We help them set up some of the things we need for the power that comes out. Then we do their further training and use all the valuable data that we have over the last 15 years to get the best outcome.

I’ll say something else, and then I’ll pause. The stack is of great importance. With almost any model we take, whether it’s a frontier model or just a ready-made open-source model, we can get much better results for much less cost thanks to the tooling we’ve built that comes with it.

So these are some of the big investments. There are still many issues related to identification, and I’d be happy to address any of them. But let me stop there and see if you have any other comments or thoughts on this.

That was going to be one of my main questions about your investment in the lab, which I think is a reasonable question for any model-training effort: Why isn’t this just a frontier model? You’re working with OpenAI and Anthropic capabilities. These labs certainly recognize the importance of security, and they have programs for verified defenders, whether it’s Daybreak or something like Mythos access. Why isn’t that enough? You have an argument about capabilities based on using your internal data. Is there anything else that’s structurally important for an independent security vendor?

Yeah, I think there are a few things here. First of all, what the labs are creating is incredible. Where they came from and what they created—it’s incredible, isn’t it? We also want to make it accessible to our customers. The general theory is that if you have a SIEM and want to use a frontier model, great. We can use it. If you want to use our models, fantastic. Our models will be much cheaper.

Just to give you an example, for a general query, if you use the top-tier cybersecurity frontier model, it’s probably going to cost you about $10. Ours costs about 3 cents. There’s a huge difference in cost, which is important. But when you look at our model, because of how we trained it on our data, the huge difference is that we have the data. For example, there is no CrowdStrike security data on Reddit. You can create synthetic data, but it’s not really the same as what we have.

We have 15 years—well, maybe a little less, because I started 15 years ago—but over a decade of security data, particularly from our MDR, our managed detection and response service, which has all sorts of threat and vulnerability pairs and attack chains. We’ve created a huge amount of data that we use for training, so we’re getting great results. I think it’s very targeted because of the post-training. Again, it’s not going to cure cancer, but it’s going to give you great results for security. That’s number one.

Number two is the aspect of security and trust. We have all the data from our customers. We’ve been trusted with this for the past 15 years. We know how to protect it, what they want, and how they want to use it. Customers are really worried. They don’t want all this data to go to the labs. They don’t want all their confidential information to get to them. They don’t want their information trained on.

11. Near Term State of Security World

We can provide them with a way to get the trust and security they want for their own security program. This concept of AI sovereignty is very important to many of our customers. Again, we’re not saying to rule out the labs, because they have incredible technology. You can choose what you want, or you can switch between different models, but ultimately we want to give customers a choice.

I think you also talked a little bit about the separation of church and state. You want to make sure that, to some extent, there’s a neutral third party that can examine some of the models the labs provide or some of the other open-source programs. It looks like CrowdStrike could play that role as well.

I think the number one factor in security is trust: being a reliable partner. You can look at suppliers. Of course, there are some big players and a lot of security companies, but trust is built in drops and lost in buckets. Customers who are cutting costs don’t want these kinds of cowboy actions that could compromise their data.

Everyone’s trying to do their best, but all our attention is focused, right? I think focus is incredibly important in any business. Our entire focus is on security, and our entire focus is on ensuring the security of our customers. We’re not trying to solve the world’s problems in artificial intelligence and other areas. I think this is definitely an advantage for us.

If you were to think about the situation in 2 years, what do you think the state of the security world would be like? What do you think will have changed by then?

Within 2 years, you’ll have local models that run on people’s computers. You could have a combination of local models plus some sort of last-mile routing to the frontier models. You’ll have many more open-weight models. You can just see the change in tokens over the last 6 months.

What changes do you see, and what do they look like? How long do you think this will last?

12. State of Model Distillation

I communicate with IT directors almost every day. I had dinner tonight with one of them from a big bank in New York, and basically the comment was: “We went all in on artificial intelligence. We went all in—great. Spending got out of control. Token maximization stopped when people got big bills.”

I said, “Well, what do you think about this? What are you going to do?” They said, “The models, the open-weight models, are so good right now. Even if they’re 3 months behind, that’s enough for what we need.”

One of the clients knows I’m passionate about Formula 1, and they said that any Formula 1 activity that we need for the bank would go to the cutting edge. For example, if it’s alpha and they’re making money, and it’s something extremely important, they’ll pay. That’s okay. But for everything else, it’s very expensive, and they’re going to create all their own open-weight models and do more post-training with them and other things. That’s how they’re going to work.

I firmly believe in this. I see this in client after client. That’s exactly how they’ll go. I think part of the challenges in the current environment is that everyone sees this, and there’s some concern at the cutting edge about what will happen.

One of the claims commonly made about open-weight models is that they’re at least partially distilled from labs. In other words, you have Chinese players open-sourcing first and effectively training their models against Anthropic, OpenAI, or other advanced models. Many of these labs are now trying to ramp up efforts against distillation, building into the model the ability not to be distilled. Do you think it’s a back-and-forth where the capabilities continue to work almost like offense and defense in cybersecurity? What’s your prediction for the state of distillation in 2 years, or within some reasonable timeframe?

My first comment is: What did the advanced models distill when they got all this information? We know about all these lawsuits about where the data came from and how it unfolded, right? So we start with this.

A fine liquor is distilled 3 times, you know—it’s triple-distilled.

13. Open Weight Model Importance and Cost

That’s right. So where does distillation start? I think that’s the way it is, but the labs are definitely focused on trying to prevent distillation. Obviously, when you look at the Chinese models out there, there are a lot of large-scale distillation systems jumping all over the internet, distilling this.

I think the challenge is that the genie is out of the bottle. There have already been many distillations. The models are very good, but we also have to give credit where it’s due. If you don’t have all the chips and you have to be inventive, and you still want to have top models, I think there are some innovative things happening outside the U.S. We should remember that, right?

It’s hard. I think it will be difficult to put the genie back in the bottle. They’ll do everything possible, but there are a lot of models. I think, now, you know what it’s going to look like in 2 years? I don’t know. You have some great models today. They can take a while, and you already have a lot of data packed into them. We’ll see how this unfolds. I wish I had a crystal ball on this. I don’t know.

This is interesting. It’s always hard to predict things like that, but I’m a little curious. Do you think there are structural advantages for Chinese companies to stay on the edge of at least open-source and open-weight models, or do you think American open-weight models are likely to catch up with them in the near future?

Some open-weight models are emerging. I’m sure you might be invested in them, too. I think having great American open-weight models will be very important. One of the reasons we worked with NVIDIA was that Nemotron is an open-source product, not just open weights. Having other alternatives will be very important.

Just being in technology long enough, you watch things unfold. For example, the operating system of the internet is Linux, period. There’s a reason for this. There will be an artificial intelligence operating system, and probably, in my opinion, it will be open source. When that happens and how it all unfolds, we’ll see.

14. Assessing Real Agent Threat

That doesn’t mean there’s no room for any advanced model. It doesn’t mean people won’t make a ton of money from it. The application stack around it—the harnesses, how things are connected, and how you think about the enterprise—all of that is going to be very important.

But ultimately, there’s just too much cost now. When you look at a lot of these advanced models as they become commercialized, if Chinese models are 3 months behind, many companies think, “Three months? That’s not a lot. I can do whatever I want.” And therein lies the challenge.

One of the top 5 CFOs told me, “A large technology buyer recently told me, ‘There’s no way their organization can afford to release a new model every 10 days.’” So a 3-month lag was faster than the regulated pace of their organization, for example. To a certain extent, it didn’t matter to them. I think openness is here to stay as a big piece of the puzzle.

One question I know we’re going to keep asking as we look to the future—and it’s almost impossible to answer—is about your history in security. This is the first time in my career in technology that some very smart, well-educated technologists are saying, “We need to shut down the internet. We just need to isolate everything because the internet will collapse next year, and it will be dangerous for critical systems to touch it.” To what extent do you think that’s a reasonable possibility in the future, when so many critical systems in society rely on the internet?

I mean, this is nonsense. Security—what agents can do and the risks they pose—is serious. Of course it is. This is serious business. We all need to look at this and say, “Yes, we need to make sure we lock everything down.” But is the world going to end soon because of these agents? No.

When you look at what agents do, I think there’s a kind of mythical notion of what an agent does. The agent doesn’t just broadcast itself from your computer to another. It actually does the same things as a human, only better and faster. So if you think about what companies like CrowdStrike have been doing for a long time, we’re preventing these kinds of attacks. Now they can happen faster. There may be more lower-level vulnerabilities tied together, and so on.

15. Areas to Watch in AI

It’s great for the adversary, not so great for the defender, but we’re creating technology that can withstand all these attacks. So what did AI do? It created more vulnerabilities, so there are more of them. And you have more exploits happening faster. So it’s bigger and faster. But this isn’t such a unique class of technology where the agent suddenly teleports somewhere.

I think people are excited about some new class of technology coming out that doesn’t exist. It was simply taught what we know. It hasn’t developed its own yet. So I look at this and think, “Security is going to be a driving force for AI again,” which, in my opinion, is incredibly important for the security industry.

Usually, you slow down: “We need to take our time.” You don’t allow the business to work as fast as it wants to. And now everyone is turning to CrowdStrike and asking, “How do we make AI safe so we can evolve faster?” So that’s how I look at it. AI is here to stay. I think there are some concerns, and there should be, but there’s a lot of doomerism around some of these things, and I just don’t think it’s necessarily justified.

I know you also sometimes work with companies as an angel investor and in other parts of the ecosystem. Are there any other areas you’re really excited about right now in terms of key areas of change or focus in the AI world, besides security? What do you think is really exciting that just happened or is about to happen?

I think chips were out of fashion back in the day. Of course, we know they’re all the rage these days, and you just need more of them, faster and more specialized. I think there will be some big breakthroughs in the chip world. When we think about speed, take Moore’s Law, for example. It may be a little outdated, but eventually chips came along and everything got faster and cheaper. I don’t know the trajectory of development, but I can guarantee that in the future everything will be faster and cheaper. I can see that happening.

16. Case for Defense Dominance

As specialized chips are developed to focus on certain areas, and new science develops around mathematics, there could be incremental increases in performance, less power consumption, and so on. So I think chips are important. Regarding open-weight models, I think there’s a lot of room for development, especially for models in the U.S. I think there’s going to be a lot of progress in shrinking these models and being able to run them on a phone, for example.

We all know the history. We remember the days when it took a computer room the size of a football field to launch the first space rocket. We have more power in our iPhones today than in that entire room. So this is coming, and I think this is going to be an area where you can make AI faster, cheaper, and more efficient, and bring it to desktops and phones, where you do it at the edge.

I think these are all exciting opportunities because I see us probably having a hybrid model, but you can do a lot of things locally right now. Again, look at TPU, NPU, and so on. They’re all in the systems we have today, and often they’ll be wasted, but I don’t think they’ll be wasted for much longer.

Given that it’s your job to think ambitiously about security all day, I wouldn’t call security people doomsayers, but they’re generally gloomy and quite pessimistic about the cat-and-mouse game getting better in their favor. Can you make the best argument for how, if people invest properly, build better products, and automate more of this, defense could actually dominate for the next decade?

I think it might, because the concepts aren’t necessarily new or unique, but the ability to quickly understand what the attack is doing and make sure your defense keeps up with it is going to be key. It’s a modern version of the book Hacking Exposed that I wrote many years ago, where the idea was that if you show people how an attack works, the defense is much better. The pace was much slower than today, but I think it will be important.

As an industry, making sure that’s key for us—and I really explain this to our team at CrowdStrike—is that you have to give the defender capabilities as good as or better than the attacker. If you’re even slightly better, mathematically, you’ll still have an advantage. Again, I think it’s always going to be a game of cat and mouse, but it’s important to approach it wisely.

17. Trailer

The only way to fight AI today is to use better AI. You have to use your advantages. The defender knows their systems. The defender has architectural diagrams. The defender has the code. So how do you translate all of this into something that can significantly increase your likelihood of outsmarting offensive AI?

The only tactical thing I find interesting is the idea that, over time, if it’s mostly agent-written code, you can implement a better security architecture and maybe have fewer vulnerabilities. Formal verification has always been extremely niche, right? But even if we ignore that, why is all software dangerous? One of the main reasons is that engineers are limited in number and capability, and we’re all trying to move very quickly.

But if you can scale with the speed of computation or your willingness to spend, and the agents have infinite patience, maybe over time we can get more secure software.

This is an interesting hope. I think there’s a point in the industry now where the code is probably good enough that you definitely get some benefits. But then again, the code was also trained in other areas, on things that could have problems.

18. Executive Advice

I have my own developer license. I build my own apps and so on, using Claude, and I have my own security agent that checks the code. It’s like, “Well, Claude just created the code, but now Claude, with its security application, says this code is unsafe.” Well, you just created it. Why didn’t you do it right the first time? Maybe there’s some optimization, but I think people don’t necessarily realize that even with zero vulnerabilities—which I hope the industry can achieve—there will still be breaches.

A lot of what we’re seeing today isn’t necessarily hacking; people are logging into the system. There’s such a problem with identity. But ultimately, if we can get better code, have fewer vulnerabilities, and distribute it, that would be great.

The problem is that it’s such a long tail. Take Microsoft, for example. Because of all the compatibility they want to maintain for backward compatibility, there are things that never get fixed or upgraded. This is just one important element of our ecosystem, but think about all the other software out there. You’re not going to fix everything overnight, so it’s going to be a really long tail.

You’ve been running CrowdStrike for 15 years, and before that you were an entrepreneur. Security—I’ve always loved it because, as a business, it’s both an interesting and important mission. And as a business, this problem isn’t going to go away, is it? I feel like a lot of CEOs of large public companies are feeling a little confused during this period because a lot of the value they’ve created during their tenure is now being called into question.

How do you think about staying informed, keeping your organization agile, and continuing to win? Is there any general advice on this?

Yes, I’m glad to take that on. First of all, you have to recognize that this is a different operating model and rhythm. We have thousands and thousands of people, right? If you think about some of the companies that have just started and specialize in artificial intelligence, they have a few people, and Claude is managing a lot of the work. They have agents that do all sorts of things.

It’s a new world, which is great, but you can’t be in the same mode as before, with that kind of waterfall development where things pop up after a year. It’s like you’re going bankrupt. So the first thing you need to realize is that having huge teams and long schedules isn’t going to solve the problem. You need smaller teams—squads or smaller groups focused on a problem—that can use artificial intelligence. You can’t have thousands of people showing up to solve one problem; you need a smaller team to do it.

And we did it with our Falcon Guardian product. To your point, I personally participated. I held meetings 3 days a week and participated in its development: “Change this, change that.” You need to go back to your startup roots because it takes that level of change in a larger company. There are many companies that are much bigger than us, but you just can’t be slow and plan.

So that’s number one: realize that the operating model and rhythm are changing, and you need to change with them. You have to disrupt yourself internally. Then you have to look at what your moats are. For example, if you’re an endpoint security company, that’s a fairly large moat. We are an endpoint security company, right? And we have data on everyone for many years.

By the way, we’re a network-effect data creator. Our sensors—we call them security agents, not to confuse them with AI agents—are software agents that run on endpoints. We create approximately 7 trillion events every day. If you’re constantly creating this new stream of data, it gives you an advantage, because the lifeblood of any AI system is the data you feed it.

19. Shifts Create Opportunity

The question for us is always: What are our moats? How do we look at this? How do we make sure we disrupt existing players? We disrupted Symantec, McAfee, and all these other players. Most of them have disappeared, changed, or been bought. You never want to be in a situation where someone is in your way, so we’re always thinking: How do we disrupt our own work? How can we do things differently? How can we think differently?

At the end of the day, if my main focus is—and I tell our team this all the time—taking care of the customer, the rest will take care of itself. The customer wants to be protected. They want to trust their supplier and partner, and ultimately, they want to do it economically. Nobody wants to spend extra money. So you have to provide the best outcome at the best price, with an experience that doesn’t require seven consoles and six people running around. You have to simplify the process.

The number one thing our customers always tell me is that it’s so easy to deploy. “Your stuff just works.” I had a meeting this morning, and someone said, “You know what? We’re merging 2 companies. They’re going to go with us. The other company didn’t have us. Your stuff is so easy to deploy.” I think that’s an important thing worth evaluating.

A lot of people are talking about job losses due to AI. That hasn’t actually happened. I think the opposite has happened so far, in terms of creating net jobs. So many questions arise about whether my children should go to school and what will happen to work in the future. What’s your opinion on this? Where would you advise people to move forward? Is this even a long-term problem?

Generally speaking, from an artificial intelligence perspective, if you look at the technological changes that have happened over the last few hundred years, every major shift has created more opportunities and more jobs. They’re different, but there are more of them. I firmly believe it will be the same here.

20. Conclusion

If you look at programming, you still need people to manage the agents. You’ll still need people to think through these architectures. It’s not just “fire it and forget it.” I know a lot of people would like to say that, but you’ll need human creativity. Cybersecurity is one of those areas I always talk about with young people and university students. It’s a great area. If you understand artificial intelligence and security, it’s an industry that’s constantly changing.

This isn’t really a solvable problem. Every vendor would like to say they’ve solved it, but it’s not a solvable problem. You need to reduce the risk and come up with the best solutions. You have to be able to protect customers, and there are a lot of things that go into that. It’s constantly evolving. In 5 years, technologies will appear that we haven’t even thought about today. If you’re a university student, you’ll be one of the people thinking about how to protect them.

I always try to give something back and think about what students are doing. Parents are always asking me, “What should my child be studying?” I think cybersecurity, technology, and artificial intelligence are fantastic industries.

This is a great note to end on. Thank you very much for that, George. Thank you for joining us.

You’re welcome. Thank you very much.

Find us on Twitter at No Priors Pod. Subscribe to our YouTube channel if you want to see our faces. Catch the show on Apple Podcasts, Spotify, or wherever you listen. This way, you’ll receive a new episode every week. Also, sign up for the email newsletter or find transcripts of each episode at no-priors.com.