EP227 AI-Native MDR: Betting on the Future of Security Operations?

Security running.

Hi there. Welcome to the Cloud Security Podcast by Google. Thanks for joining us today. Your hosts here, as ever, are myself, Timothy Peacock, the Senior PM for Detection Platform, UEBA, and Risk in Google SecOps, and Anton Chuvakin, a reformed analyst and senior staff in Google Cloud's Office of the CISO. You can find and subscribe to this podcast wherever you get your podcasts, as well as at our website. cloud.withgoogle.com/cloudsecurity/podcast. If you enjoy our content and want it delivered to you piping hot every Monday, please do hit that subscribe button. You can follow the show, argue with us, and the rest of our Cloud Security Podcast listeners all on the Google Cloud Security Community page. Anton, we have my former TLM joining us today and returning show guest joining us today. This is a special episode.

Correct. And we feel like this episode also had high expectations. I feel like for this episode, we have to deliver 10x the value.

Ah, I see what you did there. No, listeners, this is a fun episode because we get to talk about problems which are AI addressable versus AI solvable. And those are very different.

They are indeed. And I feel like since that episode, the concept is stuck in my head. And I look around and I see problems in real world, in physical world, that are AI solvable and perhaps AI addressable. And perhaps also the third category, AI can solve it but should not.

Anton, how did you get to this office building?

Oh, yes. I took a Waymo.

It was your second ever Waymo.

This was my second ever Waymo. And it was like I used my own app, and the Waymo showed up. It was quite magical.

Was that AI solvable?

I think this was AI solvable because even though I was complaining a little bit to you over chat that the Waymo is going too slowly, ultimately, this is such a future problem to have. I don't want to joke about it being a first world problem. It's the future world problem. Your Waymo is too cautious a driver. So can you please go faster?

Well, listeners, with that, let's turn things over to today's guests. And so without any further ado, I'm delighted to introduce, returning to the show, our good friend Eric and my former colleague Venkata. Eric Foster is the CEO of 10xAI, and Venkata is the CTO of 10x.ai. There's a bit of disagreement, listeners, on whether we have not in there or not. And so I want to kick things off with what is hopefully a softball. It's not going to sound like a softball. Why is your AI-powered MDR special? And why did you build this from scratch rather than say, like, why is one built from scratch going to win rather than like an existing MDR that just adopts AI?

It's a great question. I think we'll start right there with the obvious pandering. So, there's a couple of really great questions wrapped up in that. So, one, we're very heavily partnered with Google with our MDR. We believe that Google is going to win the security platform wars. We believe the hyperscalers have a tremendous advantage at this. You own your own compute, you own your own storage, you also own your own AI stack. So, one of the things that I would tell you to answer that, Tim, is we're not like training our own foundational models.

Mm-hmm.

We took money from Andreessen Horowitz, which we're leveraging to build the next great cybersecurity services company powered by AI. We believe that right now, artificial intelligence is going to disrupt the services industry in general, replacing management consultants, real estate, legal.

Replacing management consultants, I believe that part.

Oh, absolutely. I mean, we won't throw anybody specifically under the bus, but—

All my McKinsey friends, this one's for you.

You know, literally down to the name of the company, right? We believe that a 10x, taken very directly from your great—

Original stock paper from 2021.

Absolutely. The 10x stock is literally where we got the name of the company was the idea of the 10x stock.

I just wanted you to hear explicitly. So my ego gets some air.

Listeners, listeners, I need space from Anton's ego. It's bad. I'm getting smushed here.

This truly visionary white paper. Industry luminary.

All right. All right.

And Anton. But, you know, we believe truly that a 10x, a true order of magnitude shift is capable. Well, so why are we winning versus the legacy MSSP MDR guys? Well, one of the fundamental principles that's very well known in Silicon Valley, the innovator's dilemma.

Yeah.

You've built a great business as a legacy MSSP. You have people and process and technology that, quite frankly, if you could, you should tear down entirely in the age of artificial intelligence. And, you know, again, with all love and respect to your McKinsey colleagues, it's tough for them to disrupt their legacy business and retool everything for AI. It really is one of our insane advantages. And we've even been caught talking about this, of how amazing the Google engineering capabilities are, but the things that we've been able to accomplish with a very small engineering team. But then I would also say, back to my first point, we're not displacing, we're not building something that Google can build. We're building the things that Google is choosing not to build for a variety of We're augmenting the SecOps stack. We're leveraging Mandiant. We're leveraging Gemini for security. We're leveraging SecOps. We're leveraging SecOps SOAR, all of that. And then we're just augmenting components of it.

I love that.

And I think that it's almost clear to me that the main sense of an argument is that if you have a classic MDR, MSSPs, when you have people watching the screens since, you know, the late '90s, and then you say, oh, actually, we're just going to do AI. I can see how this chasm may not be— they would try to jump the chasm in 5 jumps and would not work.

Looks like you only got one jump over it.

Exactly.

It's a good metaphor.

Yeah. But I think that— I don't necessarily think that anything AI-native always wins over not AI-native, but in this case, it does make sense because you don't have this whole people replacement mindset. You have, we're building something new that works better compared to the old process. I would buy the argument.

Yes. Totally.

Even on a tech side, obviously we're building on top of Google StackOps, which is amazing at storing exabyte data and doing a lot of native functionality in it. Even from an AI stack, we are building on top of Gemini. If you compare from a cost point of view to go run an AI-native operation, it's like night and day. Gemini is around 12 cents for a million tokens versus some of the competitors are on the order of tens of dollars per million tokens, basically.

Isn't that amazing?

It is wonderful. And I tell you, my COGS conversations with Eric and my CFO are very easy. I'm like, 12 cents.

That's it.

I want to briefly—

You hear that, listeners? Gemini makes the COGS conversation easy. This is great.

I want to briefly ask a side question to this. It's kind of logical and clear why this would beat the classic MSSP MDR trying to adopt MDR, but there are other AI-powered modern MDRs people are trying to launch. And what is your stated advantage apart from using Google tech, which is going to count, which I think counts?

Sure. And honestly, that is one big component is, you know, as you know, Anton, as you know, Tim, we literally pioneered the Google security relationship from a security services side. There's essentially nobody better, more experienced on the planet at building, delivering services in the Google security ecosystem than us and our team. And so bringing that expertise to the table, being able to help our customers achieve all of the value out of their investment in Google Cloud and in Google Cloud Security is an important part of our differentiator. But 2, there are a lot of people trying to tackle AI security in general, right? In fact, you know, man, I made the mistake of walking the floor yesterday.

What did you see?

Not to throw out, not to cast any specific shade. I saw so many things that I can't talk about. But, like, not to cast specific shade, I'm not gonna name and shame here, but Very fancy security vendor who has, you know, we're the leaders in AI security operations. And it's like, man, you guys at best are a remote access tool for threat actors these days. Like, that's your main value proposition. Like, you are not leading AI security. You've bolted this on at best. And so, we're really coming truly digital native, AI native, the whole foundation of the company. Venkata, before we were able to recruit him away from Google, was literally in the AI division inside of Google Cloud Security and, you know, has that tremendous real-world expertise. I'd spent the last several years of my career in the AI endpoint space. Like, you know, we have this very practical, grounded application of AI. But then the big difference besides the— this isn't just sprinkled on, this is real use of artificial intelligence. We've got the background in the services industry. So, so many of the people that are attacking this are attacking it from a pure software perspective. And we believe—

I think I like that argument quite a bit.

And I think the next question is kind of about humans versus AI.

Because ultimately, you are an MDR, which implies that you—

and they also mentioned service quite a few times.

So, today, there's sort of a balance in your mind and in your company between this is what the machines do, this is what humans do. So, kind of like 2 sub-questions to this. Well, how does it look like today? And how do you think it would change? Because do you think that even within your framework, would machines do more? Would humans do some things forever? Like, tell us more about this whole human versus machines balance today and in the future.

Give us the giant idea.

You got it. Take a little bit of—

so I think from, again, our bottom line tagline is like human-led, machine-driven, basically.

Right.

So we believe the toughest problems in cybersecurity still need humans at the helm, making the tough decisions that only like a human can make based on years and years and years of experienced. We believe AI will absolutely make them much more efficient. We believe AI is going to be used to augment them a lot more, right? So, for example, what we do from a technology point of view is— Tim, you know this, working closely on some of the things— biggest problem in getting anything like machine learning, AI, is the label dataset, right? Saying, hey, we'll spend quarters, years maybe even, to come up with a proper label dataset, basically, right? Mm-hmm. So, what we are doing on our side is we're actually more Like using computer vision, right? And say, hey, you're working at a case, a real customer. Let us like basically observe with like computer vision what you're doing. So you basically go look into a threat intelligence thing, then you go basically CrowdStrike and do an isolate in the host or something like that. And we translate all of that using computer vision models and then basically bring it back to like the LLM to say, hey, here is the learning from the analyst.

But it has to be a good analyst.

Absolutely.

Because there was an actual startup that failed by trying to learn from really incompetent analysts using supervised ML and then transfer learning to others. So it was kind of like an interesting failure mode, which I think you're avoiding by observing the top-shelf people who know what they're doing, including some things that maybe AI wouldn't learn from just reading blogs.

It's one of the insane advantages we have. We're able to hire even better people than usual in building a cybersecurity services company because we can truly achieve this vision of the 10x SOC if we can hire 10x engineers and 10x security people. already, you know, because I'm a fundamental believer in that concept of the 10x developer, right? It's not a myth. The people who are amazing at creative and problem-solving work are an order of magnitude better. And then if you take those magnitude people, put them in one place with a shared vision, and give them the most advanced technology on the planet to give them another order of magnitude, you know, we should have been 100x.

Yeah, shouldn't we be 100x?

Yeah, yeah, we should like probably not, uh, yeah, at this point we're like, we sound like markets.

I feel like this is 5 Minute Apps. You could always have 4 Minute Apps. So, I want to drill in on this human thing for a minute and ask, what do you think today are humans uniquely good at, and where will that remain in the future? You know, 5 years from now, what are your people still going to be doing?

Yeah, that's a great question. I don't think humans are going to be out of the loop in making primary decisions in cybersecurity for the next 2 decades. Like, I'll go on record saying—

20 years?

I think artificial intelligence is Like, security is an AI-addressable problem. It's not an AI-solvable problem. Cybersecurity in general—

We should probably freeze for a moment and say, this is hugely important. This is AI-addressable versus AI-solvable. This is an important bit and probably paid for the ticket for whoever's buying tickets to our podcast. This matters quite a bit.

The tickets are free.

Yeah.

And I mean, don't get me wrong, I love Neuromancer. I'd love to see the AI models fighting each other and winning. Absolutely. But I just, I don't think that's going to be reality. I think so before moving to the AI endpoint space, I was fortunate enough to do a stint in defense tech. And even with defense tech, with, you know, some of the leading edge of artificial intelligence and automation and robotics, they're not trying to build Terminators. They're not trying to build these autonomous robots that are going to go wipe everybody out there. What they're doing is taking what used to be, and I think this is such an important analogy. For cyber that I'm making, you know, it used to take a crew of something like 35 people to fly a fighter jet. You know, it's not just one pilot, it's not just the wizzo in the back seat, it's the— this whole team of people. The new model for the Air Force, for example, is how one pilot can pilot 1,000 drones.

Wow.

And so, it's that true couple order of magnitude shift. We believe that's the same fundamental model for the future of cybersecurity. It's about having a human who's really, you know, computers are great at pattern matching. They're great at making connections and data that you might not be able to see or visualize. But we especially don't think the customers, the organizations are ready to turn over the controls of the launch, you know, command to the AI, to the robots.

Right.

They want a skilled human in the loop that they can hold accountable, that they can train, that can make the primary decisions. And so that's what we've spent a lot of our time on focusing on is how do we give those analysts better data, better context, and then how do we automate their decisions? So when I go, hey, this is ransomware, run the ransomware playbook, like here are the 10 steps that it's going to go. I can click a button, computer vision agents can spin up and computer use models spin up and do the things that I would've had to normally do.

And it's evident from our initial conversations with some of our customers right now, they don't want AI to just make the decisions for them, right? They want humans to make those decisions, but AI to assist them in making those decisions much more quickly.

Yes. And I think that comes up in a trust question because we do hear the sort of noises and concerns with agents missing things because of insufficient context, insufficient data access, or data being in some human heads, which agents obviously cannot access. So, I think I like the model with the balancing between humans and machines, but are there buyer concerns about agents missing things? And how do you deal with this? supervise what the machines do. Like, to me, this is like a building trust question. And that to me is kind of unresolved for many of the AI-only tech approaches, which makes me nervous, not because I'm like an old guy or whatever, but because I definitely see things being missed.

I also really like the subtle reference to Stanislav Petrov in that previous answer. So, shout out to that guy for saving the world real quick. But how are you thinking about this?

Yeah, I mean, we think AI can do amazing things for us. It can make our people more efficient and more effective, but we think fundamentally that those humans are going to make the critical decisions and the technology can come behind and help them.

And from a quality point of view, that's kind of like where you were going with it. There's no silver lining here. It's going to take time. That's also the reason why we believe it's going to take us a good chunk of time to get to like fully solving this problem to some extent as well. The key is we still rely on this extremely efficient senior analyst from our side where the RLHF loop really kicks in and says, hey, we think this is the cluster here, right?

Mm-hmm.

Do you agree? Do you not agree? And that's the advantage of having humans behind an AI-native solution to give you that absolute reinforcement learning right at the time of solving Yeah, that makes sense.

Part of what we've been able to do, and what I think any company that wants to take advantage of artificial intelligence in their security operations, I mean, it's building human reinforcement in at every layer, you know, so that every decision that is made, you're able to capture that learning from a human, feed it back into the model, feed it back into the training, and go from there.

So before we go to closing, I want to ask a possibly harsh question, and we are still at RSA and we're going to be walking the show floor. So, we don't want to annoy— I'm going to annoy a vendor right now. No, vendors, not by name. And then they can beat us up in theory when we walk. But I'm still going to do it. So, like, but does it mean that when somebody says, here's the box, it's AI SOC, it has no humans and it replaces a human SOC, what would you tell the customer who tends to believe in that? Like, you and me know that it's actually completely false today and in the near future. It's just flat-out false. It won't work in the real world. But, like, what do you tell the customer with some customer empathy in mind? Oh, of course. Because without using the words like idiot and stupid and dumb-dumb, basically, like, what is an empathetic answer? What is an empathetic answer?

I hope they make the right decision.

That is a great question.

I'm afraid a little bit because we're gonna go and there are 30 AI SOC vendors with baseball bats maybe waiting for us.

I don't disagree at all. And in fact, I think there's a real fine line here of trying to solve problems from a theoretical perspective and from— Yes.

Demo ready, real world not ready.

Where does the rubber meet the road? And it's part of our advantage having built and run some of the largest SOCs on the planet and built and scaled this. We understand what it's like to do cyber defense at scale. With all respect to some of those other vendors who've just literally come from the theoretical side of this, Of like, yeah, there's some interesting problems to solve here, but like, how does this actually get implemented into a real organization? How are they going to make the risk management decisions, even contractual decisions, to put this software into place that we bring a lot of that advantage? I actually think a lot of those AI SOC vendors are going to end up trying to pivot into the services space.

Yeah.

They're going to realize that it's really hard to go to market with a novel software solution that's trying to be so much intelligence, and they're going to Realize probably too late the advantage that we have coming from that human-driven perspective where we can train on analysts, we can train on their activity. We've got those skilled people at the home making those decisions, but then feeding back into the model. And then I think their investors are going to find that it's hard to take a software company and pivot it into services, especially if you've never done the services perspective.

Yes.

So that's part of why a16z made the bet on us is coming from that background.

Yep.

Of leading services companies bringing in the digital era.

Well, I hate to do this, but I have to. Traditional closing questions.

Yeah.

One, one tip to help people adopt AI in their SOC, and 2, recommended reading.

I love it. Let me take both of those, and then you should do the same.

You each get a turn, Tom.

Yeah.

You go for it.

I love it. So my first real tip, and this is not pandering, integrating AI into your SOC, you should absolutely, as we touched on earlier, You've got to look at Gemini just from a cost perspective. If you're not using Gemini in your security operations, you are probably making a costing decision unless you're just, you know, and there are a few of you who are completely price insensitive for whom you have unlimited budget. And I mean that. There are a few organizations that are listening.

Yeah.

Unlimited budget, okay, do whatever you want to do. But there is no more efficient model that we, and we've tested them all, we've used them all, we've built our, you know, a lot of people, You build a model router, you go use a lot of different things, and we're doing that as well. But like Gemini, we keep coming back to not only some of the best performance, but the best performance for the money. So, not pandering, very just literal advice to you listening out there.

Love that.

Recommended reading. So, I'm actually gonna give just a tiny variant of that question, which is I'm gonna recommend a podcast slash online reading slash everything else. So, we all in cybersecurity, We Hack Health movement. So, they've got a podcast, they've got a website, they've got some other stuff. We in cybersecurity, quite frankly, we don't do great at taking care of ourselves. It's a high-stress position. I mean, you're literally under attack all day. You've got cortisol. You're literally activating fight-flight-freeze in your day job. And then, we all also tend to be very sedentary. We're behind a keyboard. We're sitting right now doing all this. We Hack Health is a cybersecurity-focused fitness Basically collective. What?

Yeah, that is so niche.

That is so niche. That is, but that's cool also.

It's niche, but it's also cool.

But, and I mean, I literally will point out, so it started by Dave Kennedy and then Ben Cannon came together. And I mean, Dave Kennedy literally saved my life with this. Like, and I've thanked him publicly for it. Like, changed my whole health paradigm. I mean, Anton, who's known me for a while, I used to be literally 400 pounds. And no small part, the WeHackHealth movement has changed my life. For the better, my family. But telling all of you guys, invest in yourself, like to be a better cybersecurity professional, to be a better technology professional, be better for your family. Your health matters a lot, and that movement would be my recommendation. We hack health.

That's awesome. Thank you for that, Pankaj.

I said, like I say, like when you are solving a thing, like go solve a real problem with AI, and that's where you'll actually see the difference between, hey, is this AI addressable? Or is this like AI, like fully solvable kind of difference, right? So that's kind of like where even when I started like playing with like some of the AI technology like a year and a half or so ago, the demos that are out there are so flashy and so enticing. You go and put it into use, you see where the LLMs do good, you see where the LLMs are like struggling to some extent, right?

Yeah.

So focus on the outcomes, go solve a real problem, and then you can actually have a very opinionated opinion on like, hey, Can software just solve your problem in that, right? Nick's session was great on WeHack Health. He actually got me into it, which I'm going to start introducing you guys to, right? But I'll do a tech one. There is this fantastic 2-hour video by Andrej Karpathy, who is basically an OpenAI co-founder, and he used to run Tesla AI, I think, before that.

Yep.

He talks about, at depth, how to prompt LLMs correctly, how the LLMs work. It's like 2.5 hours of your best spent time if you ever listen. It's on YouTube. I think Andre started doing YouTube stuff like about 3 or 4 months ago. Please go listen to it. It taught me so many things. I thought I knew a little bit about this space, but hearing Andre taught me so many more things there.

Amazing. Eric, Venkata, thank you so much for joining us today.

Thank you, guys.

Thank you.

Thank you.

And now we are out of time. Thank you very much for listening and of course for subscribing. Please subscribe so you can get new episodes piping hot And if you love our content, please drop us a review on your platform of choice.

You can find this podcast on YouTube, Apple Podcasts, Spotify, or wherever you get your podcasts. Also, you can find us on our website, cloud.withgoogle.com/cloudsecurity/podcast. You can argue with us on the Google Cloud Security Community site, googlecloudcommunity.com.

You can follow us on X, x.com/cloudsecpodcast. Tweet at us, email us, argue with us, and if we like or hate what we hear, We can invite you to the next episode.

See you on the next episode of the Cloud Security Podcast by Google.

Security.