Rapid Response — Why everything is breaking: outages, AI, and the new cyber math, with CrowdStrike’s George Kurtz
from Masters of Scale ·
39m 9s
A surge in AI-driven cyber threats and rapid vulnerability patching has led to widespread system outages, not from AI itself, but from the overwhelming speed and scale of automated attacks. George Kurtz of CrowdStrike explains that the surge in vulnerabilities—25% of all ever-created flaws discovered in just the last year—has overwhelmed patching efforts, leading to cascading failures in systems like Microsoft, Apple, and air traffic networks. These outages stem primarily from unintended changes in complex cloud infrastructures, not from AI running amok. However, AI agents now enable attackers to find, exploit, and automate attacks at unprecedented speed, even without deep technical knowledge. This has democratized cyber threats, collapsing the traditional threat pyramid where hacktivists now match the sophistication of nation-states. As a result, smaller companies face disproportionate risks due to underpreparedness. To respond, organizations must adopt proactive risk management, prioritize critical systems, and consider service shutdowns to avoid breaches. A new AI Kill Switch provides granular control to halt rogue agents, while the future of cybersecurity lies in building proprietary, frontier-level defensive models that stay ahead of offensive AI. Ultimately, the shift is from security to safety—allowing safe AI integration while maintaining control. As the industry races to keep up, resilience, speed, and human oversight remain essential.
As you move faster and bring more AI agents into the mix, keeping track of what's actually
happening gets harder.
That's where Jira comes in.
It taps into Atlassian's teamwork graph, which connects work across Jira, the tools and apps
your team relies on most, and your AI agents.
So your agents work from the same picture your team does, and they're never left guessing.
Plus, you can track their progress right alongside your teams, so you can build with clarity
and ship with confidence.
Jira by Atlassian gives teams the context, visibility, and control to move work forward.
Try it free at jira.com.
That's J-I-R-A dot com.
Are the agents doing something that they shouldn't be doing, and they're multiplying, and they're
attacking other systems, or they're bringing down systems internally?
You know, if you're looking at everything that it's doing, you may want to hit the kill
switch.
But at the end of the day, the genie is out of the bottle.
Like, you can slow whatever you want down on the frontier.
They should have great safety.
They should have great controls.
They should have guardrails.
They should have great sandbox.
They need to do all that.
But just slowing them down is not going to slow the rest of the world down.
That's George Kurtz, CEO of cybersecurity firm CrowdStrike.
I wanted to talk to George this week amid a wave of cyber outages and rising fears of
AI run amok.
George has an insight.
From an insider's perspective, from hacker capabilities to the big AI labs, in this episode,
he talks about what's generating the recent troubles, what's likely ahead, and what we
can all do about it.
So let's get to it.
I'm Bob Safian, and this is Rapid Response.
I'm Bob Safian.
I'm here with George Kurtz, co-founder and CEO of CrowdStrike.
George, thanks for being here.
Good to see you, Bob.
So what is going on?
I mean, over the past week, we've seen Microsoft systems go down, air traffic systems go down,
like JPMorgan Chase and Apple have system problems.
No one seems to know exactly what's to blame, which gets, of course, the AI conspiracy theories
going everywhere.
Is all this just a coincidence?
You know, Bob, my phone has blown up over the last week of, you know, what's happening?
Is the world coming to an end?
And I think a few of my friends and colleagues might have their tinfoil hats.
I mean, you know, people think the world is ending here just because of everything that's
happening and going, is this really a coincidence?
And the question that you pose is, is it AI?
What is it?
And the reality is, it's the bill for the mythos moment.
Nine months ago, when mythos came out, it really set into motion everyone looking for
these vulnerabilities, right?
Not just with mythos.
Obviously, you've got sort of daybreak models, but frontier models that are capable of finding
all these vulnerabilities.
The whole industry has been racing towards find as many vulnerabilities as you can, create
as many patches as you can.
And that has taken place over the last number of months.
But what we've done now is we've hit the sound barrier of actually being able to roll
out all these patches.
So when you look at what's really taking place and you kind of look at all the issues that
were found in Black October, you've got darker November.
This is like the PHP code and now Java that people.
They're sort of scrambling to find and plug these gaps in.
Yeah, exactly.
I think the Java one is actually going to be worse.
There's already working exploits for all of that.
But what people are doing now is they're rushing to roll out all of these patches and they're
making the decision breach or break, breach or break.
So that is, are we going to get breached because we can't patch fast enough?
Or are we just going to break something and hopefully fix it if we roll it out?
And that is a. That's kind of a new norm now because the industry can't keep up, as I said, to hit
the sound barrier of patching.
And that's, we've got to find our way and our sea legs a bit in terms of how we actually
create these patches and get them rolled out without breaking things.
And as an industry, we've never had to move this quick.
And so in summary, what you're seeing, I think, is the culmination of all these
vulnerabilities that have been found, all of these now rushed patches that have to
take place.
And it's, technology's messy, things happen, and we're going to get better and better as
an industry to deal with this sort of AI speed that we're dealing with.
We're creating these bugs because we're introducing things in some ways, I don't want to say too
fast, but at a speed that we're not able to test everything before it's out.
The number one thing that actually creates outages is change.
Like things work and then somebody makes a change somewhere or it's an unintended change
or. It's one small item, you know, on the router, and then all of a sudden, you know, things stop working.
So that by far is the number one issue of why you have an outage.
And unfortunately, when you look at how complicated these large cloud infrastructures are,
you know, one little change can have a cascading effect.
So, you know, when you look at the companies that are impacted, everyone's trying to do the right things.
But because of technology, how complicated it is and how fast you have to move,
all of the unintended consequences may not be known.
Like if it's just a cloud infrastructure, you know, you've got these sort of templates
that you can update and the next time the system comes up, it's fixed, right?
That is the easier part.
When you think about all the legacy systems that are out there and you think about like
even, you know, banking type system, there's a lot of legacy type code that's out there
and it's not as easy just to create a fix while everything is in midstream.
And again, as an industry, we'll learn from that.
You've talked about clarity in these sort of outages.
You were very upfront when there was a cloud strike update in 2024
that led to some Windows troubles and some grounded flights.
But it can take time to get that clarity, can't it?
I mean, do you always know how and what created these outages?
Sometimes you never know.
It's difficult.
We went through a big one and we had our part fixed in about 72 minutes,
but you had interactions.
We had interactions with the Windows operating system that caused, you know, longer delays.
When things stop working, everyone's trying to figure out, well, what caused it?
They can see the impact of it, but they got to work back from what actually caused it.
And then really you got to start looking at, well, what changes were made?
And that's really where you start to see the issues.
You got to go back and go, okay, there was, you know, a router in Arkansas or something
that somebody put in, you know, with the best intentions and it should have been localized.
But all of a sudden it started broadcasting out.
And had other impacts.
So it's not always clear.
And when things are melting down, you get on these 24-hour bridges and people are trying
to figure out what actually happened.
It's a bit of fog of war.
As you describe it and you go through these examples, the air traffic one is a,
seems a particularly fragile one.
Is the system fragile?
I think technology is fragile.
Whether you're, you know, an Apple ecosystem, Android, you use Microsoft, you know, Google,
whatever you're using, right?
Every day people use this and they, they sort of, you know, they're, you know, they're
sort of just expected to work and they should.
But behind the scenes, when you actually look at how complicated these systems are, how
they operate, the interaction with hardware and software, you know, disks get full, memory
gets overloaded.
Like you would think that all this has been worked out, but there's a lot of just making
stuff work in the background.
And the technology folks, the men and women have to keep this running.
Literally every day they wake up and they, they're just like, okay, what are we going
to have to deal with?
We're going to have to deal with this today.
These issues of the complications within, within the systems have been there for a long
time.
We've had other kinds of cyber troubles recently, open AI's agents breaking into Hugging Face
and like French supermarkets hit by hackers, shutting down refrigeration systems, which
seem like a different order of things.
I mean, you've been in this cyber game a long time.
Any of these episodes like an aha moment for you of us sort of crossing a threshold?
When you look at what we're seeing today, it's something I talked about probably two
years ago, the automation now of the, the, the weaponization of vulnerabilities and patches.
And what I mean by that is, you know, you kind of look at the latest patch Tuesday from
Microsoft.
It was a thousand plus vulnerabilities that were fixed.
Right.
And, and again, this, this is spiked.
I mean, just over the last year, I think we've had 25% of, of all the vulnerabilities ever
created were found in the last year.
Okay.
Just to give you an idea, an order of magnitude of what we have to deal with.
So the, you know, all these patches come out and now what we're seeing is the automation
by AI agents and threat actors, where they will take a patch and they will compare what
was patched versus what wasn't patched.
And they immediately know where the flaw was.
And then they automate the vulnerability exploitation of that.
Within an hour, you can have all that done where that would take a nation state to be
able to do that maybe weeks.
So the fact that we're starting to see these issues.
Uh, come up as quickly as we are, I'm not surprised, but when you think about the refrigeration
systems, right, this was a big one in Europe.
Um, what does AI allow you to do?
There's probably a handful of people relatively speaking on, on planet earth that understand
how to program these, these refrigeration systems.
But when you actually take an AI agent and you go like, Hey, I don't know much about
the refrigeration system.
You figure out how to program it.
It will figure that out.
You literally, this is how all the TV
got hacked. You follow the Android TVs? I mean, they pointed, clawed at it and basically said,
you know, figure it out how it all works. And then you get the exploitation and then you have
like all the follow-on activity. So what I'm saying is you don't have to be an expert in any
of these technologies now, like how the refrigeration systems work. You point an AI agent
at it and then you start, once you understand how to communicate with it, then you can build all the
exploitation around it. And typically they're underprotected because it's a refrigerator.
Nobody really thought about security when they were building it.
I mean, there's some folks who say like, at this point, attackers aren't doing necessarily
fundamentally new things. They're just doing the things they used to be able to do at greater scale
and greater speed. But as you describe it, it sort of sounds like new doors are already open.
They're already open, but let me at least try to categorize the threat actors.
So the way I look at it is if you just think about a pyramid in your head, at the base of the pyramid,
you have hacktivists. So you have the most number of people with the least level of sophistication.
The next part of the pyramid, the middle section is e-crime. So you've got more sophistication,
but less of them. And at the top, you've got nation states. So the least number,
but the most sophisticated. And we used to call that the apex predator. The new apex predator
is the agent state. So now you've got the agents that sit across every, all of three of those
layers. And what you see is the pyramid is now collapsed. And what I mean by that is the
hacktivist has the same capabilities as the nation state. Using AI, you have the same capabilities.
You can create a very, very sophisticated attack just by talking to the AI agent.
So what that means now is you've got a higher number of attacks, like greater volume,
and you've got a greater level of sophistication just driven by automation. And that is the scary
part of where we are today, is keeping up with the volume. But it's democratized destruction.
Anybody with an AI agent, I mean, some of the frontier models will refuse what you want to do.
But if you take an obliterated model, which you can download from Hugging Face,
and you plug it into your Mac or a Spark, DGX or something, you can run the most sophisticated
attacks without any refusals. And that is the state of the state for where we are today.
All these changes, these AI fears and vulnerabilities,
they're good for your business, aren't they?
Technology change is good for our business and for security. And what I mean by that is
security has always paralleled the slope of the technology curve, because it used to be simple,
and now it's super complicated up to the right. And every time there's a change in technology,
whether it's a new technology or a new technology, it's a new technology. And so,
whether that is cloud or whether that's mobile or whether that's AI or went from mainframe to
client server, all these sort of things, you needed technology, you needed security to go
with it. There's a shift now from security, which I've done over the last 30 years, keep everyone
out, right? Security business, keep everyone out. Two, safety, allow everyone to use technology.
So, one was keep everybody out. The next one is, how do you safely allow,
the use of AI? In other words, how do you safely allow your own team to go out
and come back and not bring things with them? Exactly. And this is security and safety. They
sound similar, but they're vastly different. And the safety element now opens up a much broader
addressable market for us. How much of the game is becoming like,
my AI is better than their AI? I mean, because you have to stay ahead of the attacks.
It's, you know, the old mad cartoon, spy versus spy. It's really what it is, right? Who's going
to stay ahead? And what we've done is we've actually formed a super intelligent cybersecurity
lab to create essentially frontier caliber models for security, both on offense and defense.
And the offensive models continually make the defensive models better. Now, why is that important?
Because right now, the offense, the bad guys, the bad threat actors have an advantage over
the defenders because the defenders didn't have models capable to deal with the attacks they were
seeing. Now, a good example of that, you may say, well, how can that be? Let's take hugging face as
an example. People look at this and say, wow, the aha moment was you had autonomous agents
breaking out of a sandbox, breaking into another company. They did it to pass the
test, but it was done end to end. Sure, that's novel and unique. But the aha moment was the
fact that the defenders didn't have the same frontier capabilities in their models as the
offense. And what they had to do is they had to go out to a Chinese model and they had to figure
out all these attacks, what was behind it to make sense of the fact that it was actually coming from
these AI agents. So the defense didn't have the same,
capabilities as the offense because the models refused. They went to frontier and it just said no.
So that's why in today's world, the only way to solve this problem
is to have defensive models that are better than the offense.
I could imagine that in the old world, if I were a hacker at any stage of the pyramid there,
I would probably target my efforts at bigger places because the return was bigger. It was
the effort. But now it seems like it's easier to spread those attacks anywhere. And I guess
that probably means that smaller enterprises and smaller outfits are vulnerable. I mean,
maybe they were always vulnerable, but they're targeted now in a different way.
Most things in life come down to time and money. I mean, just to simplify it, right? So
how much time does it take? How much money can be gained? In this world,
you can go to bed, let your agents run. And for $50 in tokens, you come back,
the next morning, you've got a new bounty of riches of companies that the agents have broken
into. So you're exactly right where when we think about ransomware or companies being extorted,
the big multi-million dollar payoffs are certainly going to be in the large enterprises.
But what we're seeing now is because of the automation, there's a lot of money to be made
in the SMB. Small dollars add up to big dollars. I talk to people all the time and they say,
we need something like CrowdStrike because we don't think we're going to be attacked or we're
small-time. Why would we be a target? Yeah. And that is just not the case. You can make $5 million
from one company or you can have a hundred companies at $500,000. Due to math, they're
going to be underprotected compared to a large enterprise. And generally what happens, they get
knocked offline and it's not as easy for the smaller companies to recover. And it's a big mess.
And tech, they don't control, you know, cloud platforms and AI models and vendor software.
How much can you trust a partner to provide the security for you?
I think you got to look at the vendors that you're dealing with,
how much you're spending on security, what sort of certifications they have.
And then you have to then look at like the supply chain attacks, which are very prevalent now.
What I mean by that is over the last 30 plus years, we've said, keep everyone out. Don't
run untrusted software. You know, if you're going to download it, it has to be, you know,
checked and secured and those sorts of things before you run it. And we've turned that entire
model on its head. So if you're running a cloud or a codex or whatever, you know, pick something.
Generally these tools will go, Hey, I'm like, I need something open source. I'm going to go out,
download it and just run it. And these packages get infected. So we took everything that we were
doing defensively over the last 30 years. We threw it out the window and we're like,
and we said, Hey, we're just going to let our agents, which we're going to give our identity
to or worse, it steals an identity that's even greater than our own. And then it can go out to
the internet, download these packages and just run whatever it's going to run. Like that's the
scary part of what we're seeing now. I always thought of supply chain in a physical sense,
but what George is talking about is the digital supply chain where information is brought in by
AI agents. So does that mean every business needs an AI agent?
Kill switch. We'll talk about that and more after the break. Stay with us.
How to protect what you've spent years building. Creative planning where wealth works together.
Learn more at creativeplanning.com slash masters of scale. It taps into Atlassian's teamwork graph, which connects work across JIRA,
the tools and apps your team relies on most and your AI agents. So your agents work from the same
picture you're working from. So you're working from the same picture you're working from. So
your team does, and they're never left guessing. Plus, you can track their progress right alongside
your teams so you can build with clarity and ship with clarity.
confidence. We've all spent money on something that
seemed incredible but didn't live up to the hype. A kitchen gadget, a car, when it doesn't deliver,
the disappointment is real. Marketers suffer from this too, chasing metrics that look great
but don't translate into revenue. LinkedIn calls that bull spend. Invest in what matters and reach
the right buyers with LinkedIn ads. You can target by company, industry, job title, and more.
LinkedIn ads also generates the highest ROAS of all major ad networks. Spend $250 on your first
campaign on LinkedIn ads and get a $250 credit for the next one. Just go to linkedin.com. That's
linkedin.com.
Terms and conditions apply.
Before the break, CrowdStrike's George Kurtz explained how a wave of efforts to patch
cybersecurity vulnerabilities has ironically precipitated a flood of outages. Now he talks
about a new product called the AI Kill Switch, why everyone should hug their IT people, and more.
Let's jump back in. You recently launched a product called the AI Kill Switch, which is
designed to stop AI kill switch. And it's a product that's been around for a long time.
It's a product that's been around for a long time. To stop agents, rogue agents. How do you know which scenarios would prompt hitting the kill switch?
If you think about industry, if there was a manufacturing, a big machine that had a problem,
there's always the red button. Like, you know, if you walk around any plant, you'll see the big red
button, right? It's like, stop that thing because, you know, it could be causing bodily harm.
And it's the same scenario. Are the agents doing something that they shouldn't be doing,
and they're multiplying, and they're attacking other systems, or they're bringing down systems,
internally? You want to be able to hit the button. When you kind of look at the hugging face
example, you know, if you're looking at everything that it's doing, even if in an automated way,
you can see that it broke out and what it's doing, and you may want to hit the kill switch. And
the challenge with the kill switch is there's just blunt force, like, it's almost like pulling the
plug. Right. It's pulling the plug on everything. On everything. Or do you go like, hey, this agent
and its sub-agents need to be basically killed. And those are the things that we continue to
refine.
So you can get granular and not just pulling the plug for an entire company.
Over time, your switch will get, be able to be more targeted than it is right now.
We can do it in mass, and we've got, you know, more granular controls. But over time, it gets
more and more granular down to the exact command, the exact agent, the exact sub-agent.
And those are the things that we continue to add into the product.
I wanted to ask you about the proliferation of deepfake.
How widespread should we expect that sort of deepfake capability to grow? You know, and like,
if you can't trust a voice or a face, do we have to rethink verification in some ways?
It's a real problem. You know, I can change my appearance to be anyone I want. The tools are
there. There are AI overlays. So, Bob, you can be talking to me and the AI is listening to you
and then immediately providing sort of responses back.
A good example of this is the North Koreans when they essentially are getting jobs as tech workers
and they're passing with flying colors all these technical tests. And these are the way they're
doing it. Like, they're in an interview and they pass with flying colors because it's kind of
listening and telling you how to respond. And then they're infiltrating companies and creating a
laptop farm. And then ultimately, they have access into your company. They don't have to break in.
They just, they're already logged in.
They're working there.
They're already working there. So, deep fake is a massive, massive problem today.
And authentication and identity is very, very important. You just can't take somebody's
word for it or the fact that I'm going to get on Zoom with you and it looks like Bob in your place
there. So, this week, Anthropic is going public. You work closely with both OpenAI and Anthropic.
Dario Modi at Anthropics called on the industry to slow down. There's been quite a lot of noise
about it.
Although development keeps steaming ahead. Do you give OpenAI and Anthropic advice about
speeding versus slowing? Do you worry about developing AI too fast?
Yeah. Look, I'm not sure whose advice they listen to. We worked with Anthropic to be able to take
our technology and run it in the Claude Co-Work container. So, all of that is very positive. We
have a very good relationship with them. You can slow whatever you want down on the frontier. The open weight models are three months
behind. I mean, they're good enough and you can run them without any restrictions because they
take all of the guardrails out, right? Which is called an obliterated model, right? It comes from
the term ablate or remove. So, you can go to hugging phase. You can say, just get rid of the
refusals and you can have it do whatever you want. So, they can slow down all they want. And by the
way, they should have great safety. In the spy versus spy world that you live in,
you need those frontier models to continue to be at least three months ahead or more because
you're using those models to stay ahead of the folks who are attacking. We do work with them.
We use them, but there's a reason why we created our own labs is because we're not going to just
depend on the frontier models. The thing is, they don't have all the data that we've accumulated
over the last 15 years. So, we've taken open weight models and we've post-trained them with
our own data. So, we're getting frontier or better level offensive models and defensive
models at a fraction of the cost. And then we've built our own harness. We certainly will leverage
any of the frontier models. They're doing great work.
But we're not going to just depend on what they're doing. We built our own.
And that's why we have now a massive lab ourselves that in the cyber world is operating at frontier
level capabilities. Our models are not going to cure cancer, but we want the best cyber models
out there. And that's what we're working on. What perspective do you have about what role
government plays in the evolution of this? AI labs policing themselves, you guys policing
yourself with your own.
Well, I think standards and guidelines are important. And, you know, there are other
areas for policing. I can't wait for open AI and anthropic to go public because you're held to a
much different standard. There's so many compliance requirements. There's so many
disclosure requirements. You know, with that comes more liability. So, I think there's a natural
forcing function there. But the government role is, you know, how do we maintain our lead?
In AI, SI, you know, pick your acronym. I think if you just clamp down on, you know,
we're going to have the FAA for certifying models, like that's not going to work because the rest of
the world is moving very quickly. So, what are customers saying? They're saying we want to work
with the most secure and safe providers out there. If you're reckless and your agents are doing bad
things, you're not going to be able to do that. You're not going to be able to do that. And you can't control them. And you don't give visibility and
sort of, you know, provide safety from a model lab perspective. Then I think the market is going
to see that and go in a different direction. But keep in mind, and I just want to make this
distinction, Bob, they are called labs. Okay, labs. What happens in a lab? You develop something,
it kind of works, it does great things. But when you put it in the real world, it may not actually
work as intended in the lab. What happens in the lab? You put it in the real world, it may not
which is why the security industry has now moved into the safety industry and has become a great
opportunity for folks like us. And you as a, or anyone as a lab operator, you have a choice as to
what and when you release things that you've tested in the lab into the real world. That's it.
Look, anyone can slow down. If the model is not right, if it's not aligned, if it's not getting
the output you want, if the quality control is not right, if it's not getting the output you want,
if the quality control is not there, you know what you do? You don't release it. You go back and do
another run until you get it right. And then obviously, you know, these things are
non-deterministic. They're still going to have to fill in the blanks when they get to the real
world. This is part of the problem. And, you know, just to try to simplify it a bit is the models get
trained on what it knows. Okay. And then it's almost like a kid that you raise, you give it,
morals and values and a framework for making a decision. But when it gets in the real world,
there's a scenario that's never been seen before. And it has to use that framework.
And then it has to reason and make a decision. This is how a model works. So they can do the
best training. They can put guardrails around it, but at some point it's going to get in the real
world and it's going to have to fill in the blanks of what decision it's going to make.
And how that judgment is determined. I mean, we struggle with,
that with our children, it's hard to find that
too. I mean, that's what that strange term interpretability, I guess, is about in the AI
world, is trying to figure out how it's making that judgment. Exactly. And unfortunately, it's
a bit of a black box. Part of what needs to be done is to understand the reasoning chains of
how the model is actually deciding what it's doing. That generally isn't exposed, and I think
it should be. Beyond CrowdStrike, you're an endurance race car driver, and a good one. You
won your class at Daytona and Le Mans, some other places this year. Both car racing and cyber
security involve accepting a certain amount of risk. Are there lessons that apply from one to
the other? I think there's two critical concepts that apply. One is managing risk, and two is speed.
Because when you're driving a car and you're racing, it's all about how much risk you take.
Risk passing the guy in the corner, the risk, you know, stuffing it in under braking and having an
issue. Is it a clean pass? You know, you're in traffic. How much risk do you take? Is it the
beginning of the race? Is it at the end of the race? That's security. Like, how fast do we go
patch something? Are we okay with braking something versus having a breach? There's a risk. There's a
trade-off. And then there's speed. At the end of the day, there's speed in the stopwatch. In security,
paramount. It's always been important, but it used to be like weeks before people patched.
We'll have it patched in 30 days. That is long gone, right? So the speed and the cadence that
you have to deal with this is the same thing in racing. Like, you got to move fast. If you're
in racing, the saying is, if you're not moving forward, you're moving backwards.
It's the same in security. So what should any business leader do right
now, immediately, with speed, to improve their cyber footing? And I guess, what should we be
preparing ourselves to endure as we go through this period of tech change?
Well, I think maybe first is go find your security and IT people and give them a hug,
because they are at 110% of capacity now. We talk to teams over and over, and it's not sustainable.
So, and, you know, I think it's focusing on how do you manage risk? What's most important?
How do you prioritize? They're not going to be able to patch everything as quickly as they'd like.
So understanding critical assets, understanding exploitability, so that you can prioritize what
happens. I've never seen this before, but it's happening across industry, is
companies, CISOs, CEOs are making a decision. Take something offline,
or be hacked. Like, so they're saying like, hey, we would rather take a massive service offline
than be hacked. Because the risk of being hacked is so much higher.
Exactly. So think about these big services that we depend on. We work with the largest enterprise,
and they're like, hey, we'll take down this part of the service because, like, we would rather just
take it down than have a security issue. And that would never enter the equation a year ago. But now,
they're willing to take it down. And so, I think it's important to think about that. And I think
it's important to make that decision if it's so bad. Like, when PHP came out, I had plenty of
companies go, hey, like, we may just take everything down. Because it's so bad.
And a year ago, what did they do?
A year ago was like, hey, just, you know, do your best, patch it, mitigate it, put some rules in
place, you know, keep everything running. And now, with AI agents, how fast they move, and this
concept of a swarm.
Like, they're basically saying, hey, we may just take it down. Because if the AI agents get in,
A, you have a breach, but you might be down 30 days, then, you know, a few hours of taking it
down because you got to figure out all the breaches, you know, where it all is. You have
defensive AI, which is to keep the offensive AI agents out. That's one. And two is, how do you
use security to provide safety so that you can, you know, keep the AI agents out?
You can actually exploit the power of AI. And that was one of the number one things that I've
talked to CEOs and CIOs, Chief Information Security Officers, where they've said, we need to go
faster and deploy AI for the benefits, but we can't unless it's safe and secure.
Give you a racing analogy. If you're racing a Miata, you got a lot of time to make a decision
whether, like, you know, you should pass or not. Like, it's slow motion compared to a Formula One
car.
So, you may have to make decisions faster and accept some level of breakage until the industry
gets better at understanding how to patch and keep things running in the wake of these AI attacks.
Well, George, I really appreciate this. I'm not sure if I feel calmer and more relaxed,
but I feel more informed.
Well, it was great to spend some time with you. I think on a positive note,
I am an AI optimist, and I think there's enough technology and security that can keep us safe
as a society. But with any great power comes great responsibility,
and we've just got to harness it and make sure we use it in a secure way.
Amen to that.
I want to double click on a stat George shared. 25% of all cyber vulnerabilities ever created
were identified in the last year. Companies used to try patching trouble spots,
but now they're preemptively shutting down to forestall any potential invasion of rogue AI
agents. What George is telling us is if you feel like the cybersecurity environment is in flux,
you're dead right. Still, that doesn't mean everything is about to collapse.
I'm mindful of George's car racing specialty as an endurance racer. That means he keeps the long
term in mind with each move he makes, regardless of how fast he's moving. In today's world,
that seems like a problem. But I think it's a problem. I think it's a problem. I think it's a
process we all need to get comfortable with. I'm Bob Safian. Thanks for listening.
We all know the disappointment that comes when a grand gesture falls flat. Something that seemed
impressive in the moment, but didn't deliver the results you expected. Marketing budgets too often
fall into that trap, aiming for big impressions, big reach, but without the revenue to back it up.
LinkedIn has a name for it, Bullspend. LinkedIn ads generate the highest ROAS of all major ad
networks. They target by company, industry, job title, and more. Just go to linkedin.com slash scale.
That's linkedin.com slash scale. Next, with Stripe Atlas. Start your company with incorporation, founder equity setup,
and startup legal documents all in one guided process. And once your company is formed,
which for most people happens in just one day, you're ready to begin accepting payments on day
one. More than one in four new Delaware corporations is formed with Stripe Atlas.
Learn more at stripe.com slash atlas.
Rapid Response is a Wait What original. I'm Bob Safian. Our executive producer is Eve Trow.
Our senior producer is Alex Morris, and our associate producer is Masha Makutonina.
Mixing and mastering by Aaron Bastinelli and Brian Pugh.
Our theme music is by Ryan Holiday. For more, visit rapidresponesshow.com.
Hey, listeners. Bob here. If you listen to Rapid Response on Masters of Scale,
you may be missing half the show. Because every Friday, we release a second Rapid Response
exclusively in the Rapid Response feed. The guests and topics are just as compelling and timely,
from Ford CEO to NASA's administrator to the lessons from the NASA team.
The devil wears Prada. It takes about 10 seconds to find. Just search Rapid Response wherever you
listen to podcasts and hit follow to make sure you never miss an episode. I hope to see you there.
Podcast Summary
Key Points:
Rapid AI-driven vulnerabilities and patching efforts are causing widespread outages, not due to AI itself, but because of the speed and scale of automated attacks and rushed fixes.
AI agents can now exploit system weaknesses—like refrigeration or software code—with minimal expertise, effectively democratizing cyberattacks and making them faster and more widespread.
The traditional cybersecurity model of "keeping everyone out" is being replaced by a new paradigm of "safety through controlled access," where organizations must safely integrate AI agents into workflows.
A new "AI Kill Switch" allows teams to stop rogue or harmful AI agents, offering granular control to avoid system-wide disruptions while still enabling agility.
Cyber threats are now distributed across all layers—from hacktivists to nation-states—thanks to AI, creating a collapsed threat pyramid where attackers have equal capabilities.
Smaller businesses are increasingly vulnerable due to underinvestment in security, as AI attacks can be deployed cheaply and quickly, leading to significant financial and operational damage.
Organizations must prioritize risk management, assess critical assets, and consider intentionally shutting down services rather than risking breaches from AI-driven exploits.
The future of security lies in building advanced, in-house AI models that outpace attackers, ensuring defense capabilities are not dependent on external, potentially lagging models.
Summary:
A surge in AI-driven cyber threats and rapid vulnerability patching has led to widespread system outages, not from AI itself, but from the overwhelming speed and scale of automated attacks. George Kurtz of CrowdStrike explains that the surge in vulnerabilities—25% of all ever-created flaws discovered in just the last year—has overwhelmed patching efforts, leading to cascading failures in systems like Microsoft, Apple, and air traffic networks. These outages stem primarily from unintended changes in complex cloud infrastructures, not from AI running amok.
However, AI agents now enable attackers to find, exploit, and automate attacks at unprecedented speed, even without deep technical knowledge. This has democratized cyber threats, collapsing the traditional threat pyramid where hacktivists now match the sophistication of nation-states. As a result, smaller companies face disproportionate risks due to underpreparedness.
To respond, organizations must adopt proactive risk management, prioritize critical systems, and consider service shutdowns to avoid breaches. A new AI Kill Switch provides granular control to halt rogue agents, while the future of cybersecurity lies in building proprietary, frontier-level defensive models that stay ahead of offensive AI. Ultimately, the shift is from security to safety—allowing safe AI integration while maintaining control.
As the industry races to keep up, resilience, speed, and human oversight remain essential.
FAQs
The AI Kill Switch is a tool that allows organizations to stop rogue or problematic AI agents mid-operation. It acts like a red button in a factory, enabling teams to halt agents that are causing issues, such as attacking systems or multiplying uncontrollably.
AI agents can automate the discovery and exploitation of vulnerabilities at an unprecedented speed. They analyze patches and identify flaws in seconds, whereas human attackers might take weeks. This automation allows AI to find and exploit weaknesses much faster than traditional methods.
The speed at which AI models are developed and deployed means vulnerabilities emerge faster than they can be patched. This creates a race where attackers can exploit new flaws before defenses are updated, leading to widespread outages and security breaches.
Smaller businesses often lack the resources and expertise to respond to complex threats. With AI agents capable of launching attacks at scale and low cost, even small-dollar attacks can accumulate to significant financial losses, making SMBs prime targets.
The automation of weaponization means attackers can use AI to turn patches into exploits automatically. This allows even non-experts to launch sophisticated attacks, dramatically increasing the volume and sophistication of cyber threats across all industries.
Guardrails and safety controls prevent AI agents from performing harmful actions. However, some models—especially 'obliterated' models—remove these safeguards entirely, allowing full access. This highlights the need for robust, enforceable safety mechanisms in all AI systems.
Chat with AI
Loading...
Pro features
Go deeper with this episode
Unlock creator-grade tools that turn any transcript into show notes and subtitle files.