EP172 RSA 2024: Separating AI Signal from Noise, SecOps Evolves, XDR Declines?

Security.

Hi there. Welcome to the Cloud Security Podcast by Google. Thanks for joining us today. Your hosts here in person together in the same room, myself, Tim Peacock, the Senior PM for Risk Analytics, Cloud Threat, and Google PsychOps— announced this week that that's the real name— 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.google.com/podcasts. If you enjoy our content and want it delivered to you piping hot every Monday, please do hit that subscribe button. Follow the show, argue with us and the rest of our listeners on our community page. RSA Week 2024, pretty big, huh? What do you think?

Yes. And now we are— it's still Wednesday midday when we're recording this, and we are told to provide profound insights so that our audience the world over can learn from what we learned.

Oh, God.

So it's kind of high expectations, I guess. Is this the episode where we go from banter straight to the episode, right? Because we kind of like, that's what we're doing. And what about starting from something moderately safe? Like, I don't know, what did you launch?

What did we launch? We launched Google Threat Intelligence. That's pretty cool. We brought together everything Google knows about bad guys, made it available in one place, unified our visibility, and we have unparalleled depth, breadth, currency, and expertise brought together in one place.

You sound like a marketing person, Tim. Sorry.

Once in a while I fall into that.

This unparalleled stuff is just making me nervous.

Okay. Okay. But who's got, who's got more depth than VirusTotal?

VirusTotal plus Mandiant.

Who's got more recency of responding to leading breaches than Mandiant?

Yeah, that's nobody.

Who protects more browser sessions a day than Safe Browsing?

That's probably nobody again.

Correct. So when I say unparalleled, It is literally unparalleled.

So it's not an exaggeration. So it's like one of those things when you say you 10x something and people question you and they say, here, here are my measurements. It really is 10 times better. We've seen that happen with some of the projects and stuff. So, yes.

We saw some interesting marketing today. We saw a competition on booth language between optimal, better, and done right.

Yeah. I mean, this was like, for some reason, maybe it's my past being sort of in marketing that really triggered me. Because when you show up with booth, of course, Of course, you can go buzzwords and say you're zero trust powered AI, powered by ML, powered by—

That stops ransomware.

Yes, or that stops ransomware. But you can also go the other way and say that you have the optimal tools and they're done right and they're better. But it's like still, which would you choose? Excessive buzzwords or deathly boring?

Neither.

Right, exactly. Thank you. Like neither crazy, over-the-top, silly messaging about AI-powered ML. Was it ML-powered AI? One of the two.

Both probably.

Yeah, both probably. Versus like we have an optimal—

Totally vacuous.

DLP that actually works. Actually, that's an extreme statement. DLP that actually works.

Oh, boy.

So, okay, right. Let's avoid this one. But as far as buzzwords, what did we see? AI powered by AI.

I think we saw a lot of that.

We saw a lot of AI.

A lot of AI.

A lot of AI for security and a little of securing AI.

Yeah. And we had an interesting argument with a podcast guest, a former guest. We'll leave her anonymous. But we had an argument with her about whether all this AI stuff was good or not. And I think this was an interesting line of reasoning because, listeners, what we argued was that AI for security is in a state where we don't know which use cases are going to turn out well.

Mostly GenAI in this case, right?

GenAI specifically. We don't know which GenAI use cases are going to turn out well. And so as an industry, it's not unreasonable for us to all spend some time and money and CPU cycles, GPU cycles, Figuring out which ones work. And we don't believe you can do it any other way.

Yes.

We've been talking about things that you can figure out in a basement by yourself with a book versus things you have to discover. And I would say this is— you have to discover.

Yes. And I think that this is the— you can make jokes about annoying marketing, but ultimately the industry very obviously is in the use case discovery phase for AI. And some people would try hard to go with a use case, they would market the hell out of it, and the use case would flop. And that can happen. And other people would go hard, discover a use case, and it would be fine. And that's also possible. But to me, I would decouple the use case discovery from over-the-top marketing.

Yes.

Use case discovery is very real. Everybody is realizing very quickly that summarizing and explaining things using GenAI works really well. It may not sell tickets to your GenAI show, Or, but it actually helps.

It works.

It works. The VirusTotal example, when they reversed malware using the Gemini with increased token windows, was kind of epic. To me, that's one of the magical use cases. It's niche in terms of who needs it, but it's epic in terms of impact.

Yes.

And then people who go and ask a consumer-grade AI to create detection rules are going to discover that— They have a bad time. Yes. Thank you, Tim.

Yes.

I wanted to say something more rude, but thank you for stopping me.

They're going to have a bad time. They're going to have a bad time.

But it's Bedtime TM with capital B and T.

Yes, yes, yes. I'll trademark that phrase. That's great. So we saw that. We saw not a lot of XDR. XDR is down.

I think XDR is— did you say down or dying?

Down.

Down.

The trend is down.

The trend's down, but there's still people with little— who maybe just got them emulated or something, then they're doing XDR. But ultimately, XDR, to me, if I were doing some kind of a hype cycle, But to me, XDR very clearly fell into the ditch.

Yes.

And the question is, would it emerge as something useful out of a ditch, or would it stay in the ditch where things tend to kind of—

Stay?

No, they essentially disappear out of the ditch. Oh, they decompose. Yeah, they decompose. Correct.

Got it.

But actually, there's some other old things, email security, even DLP, even some other stuff.

There were a lot of them.

A lot of them.

A lot of them and a lot of new ones. Listeners, we found several email security companies in the back corner, low rent district, emerging players area. And I was shocked that the world needed another email security company. In fact, it seems to need at least 3 more based on what we saw.

Yes. And while we can always make some kind of self-serving joke about you not needing it with Workspace and with Gmail, there's probably enough organizations who do need better email security, so they're buying—

Oh, you mean like the US State Department?

Well, that's an extreme case, but yes. Not that I want to joke. I don't know if anybody still uses Lotus Notes for emails, but even 5 years ago, there were definitely people who were in that bucket. Actually, which reminded me of another lesson that I've seen is that in our industry, when people think about the future, the past is actually very strong. I've seen the— whatever, who was it? Silence, BlackBerry, EDR. There were some SIM vendors who—

Carbon Black, Symantec was there.

Carbon Black, Broadcom, a few others that you typically associate with the past.

In the ditch.

In the ditch is too harsh to me. But the point is that these are vendors that were very strong.

20 years ago.

I think you're still being too harsh. Sorry, I'm the mellow guy today. I don't know what happened.

What's going on, listeners? This is a weird day.

Yeah. But imagine like a third-tier DLP vendor who was really cool in the mid-2000s, early 2010s. They have a booth, sizable booth perhaps. So the pull of the past is very strong. And a lot of organizations are using some things that the cloud literati would probably laugh at, but it's very real. And I think that the Innovative stuff is great and it's happening, but a lot of the industry is using that stuff and it's there and it's not going to go away. By the way, which reminds me, I had another epiphany.

Another?

Another epiphany. It's this whole platform thing versus best of breed.

Okay.

And the epiphany goes like this. There's an argument in the industry about big platforms versus individual tools.

Yes.

And I think there's no argument. And here's why there's no argument. When you enter the RSA show floor, You see 700 vendors or something.

At least.

At least. And if you look at various catalogs, IT Harvest has the longest one, probably like they count 5,000. A few others have like 2,000. The point is that very clearly, we are voting with wallets that we buy niche solutions. There is no platform. The argument does not exist. The platform side does not have a weak argument. It has none.

Because clearly, people vote with wallets to buy single But what if we discovered that the platform is getting 80% of the wallet share?

Yes, that's a good counterargument. I have not really created a counter-counterargument to that.

Oh, good.

No. Well, it doesn't mean you win the argument. It just means that it's still a good debate to have. But the fact that RSA show floor exists and the fact that people pay for all this stuff, otherwise, if people didn't pay, the booths wouldn't exist. Well, VC money and understanding, that means that We are kind of voting for best-of-breed single-purpose tools in many cases.

I think the other thing that this says to me is that there's an extreme number of niches in cybersecurity.

Yes. Some niches are clearly on the growth phase, but some niches stay nichey for years.

But that does mean they're not profitable if you want to build a 20-person security company that solves a very particular problem.

Which no large platform vendor wants to solve or can solve, or if the problem is new, maybe it's securing AI or Securing specific type of OT in a specific type of vertical.

Yeah, yeah, yeah. Securing left-handed screwdrivers only sold in New Zealand. Like, somebody does that.

That's a bit pushing it, but yes, I think that's—

But you know what I mean.

Yes, yes, yes. I mean, but some niches I think are expected to expand. You know, we spotted ASPM.

What is ASPM?

Okay. So the way I explained it in one of the blogs, I can mention my blog, right? It's the guests cannot, but I'm kind of not a guest. And so There was a point about that there are 2 gangs, the SPM gang and the DR gang.

What do they stand for?

CDR and CSPM.

What do these stand for?

Security posture management and detection and response.

Okay.

So, for example, for cloud, we have CSPM, cloud security posture management, and we have CDR.

Cloud detection and response.

Cloud detection and response.

And these are the Sharks and the Jets of the security world. They have snap fights, dance battles.

Okay. No, they're also kind of friends sometimes.

Interesting.

When we see that something— there is also the SSPM, which is for SaaS. security posture management, which does not have the SDR, funny enough, maybe not yet.

But there's also— Well, that's software-defined radio.

Yeah, it would be that. But there is application security posture management, which is emerging. What does that mean? You know, that's a good question because if it's all about cloud applications, in theory, CSPM should cover it. But there are enough massive-scale SaaS applications that are very complex, and they may justify their own application security posture management. It isn't just cloud infrastructure.

So wait, you're telling me that there's a whole field of I secure, say, SAP or Salesforce?

I don't want to name names, but there are big complex enterprise applications in the cloud or not that may justify its own posture assessment toolset.

Wow.

And perhaps its own detection response toolset because people are launching ADRs as we speak. Not to make another mention of the blog. Okay, fine, I won't. The point is that ADR may come as well as a niche.

Fascinating.

Application detection response.

We saw several vendors with those letters on their booth. So this exists. Okay. We continued a tradition this year, listeners, and we've done this at least 2 RSAs and 2 NEXTs where we went to a series of vendor booths and asked the same question.

Yes. It's about cloud. So we are now switching subjects to cloud security, our favorite. Well, one of our favorite topics.

One of my favorites. It's up there.

Yes. It's up there.

What's the question? What do we ask every year?

The question is, so one of my former colleagues, meaning a Gartner analyst, have coined the term called CNAPP. It's one of those rare and precious 5-letter acronyms that everybody loves. What does it stand for? You know, I know that what PP stands for. It's protection platform.

Yep.

CN stands for cloud native, and A stands for, shockingly, application.

Weird.

It is. But the CNAPP is kind of a duct tape bundle of stuff.

It's your Sharks and your Jets put together.

And a lot of other things.

And a lot of other things.

But the things that actually matter are these exact 2 things, CSPM and CDR. And for the last couple of years, we would come to a lot of booths of vendors who are now all CNAPP and say, hey, you do cloud security? Great. When people bring you money and they give you money to get your tool, what do they actually buy? And specifically, do they mostly buy CSPM?

Or CDR.

Or they buy CDR or something else.

Yes.

This year we added this something else option, which of course nobody picked.

Right. And we got different answers depending on who we asked.

Yes, correct. But I think the popular answer was yes. When we sell a broad and beautiful cloud security platform that matches Neil MacDonald's CNAPP vision, they ultimately mostly buy CSPM.

That's right.

And some people said that they also buy CDR at increasing percentages.

Yes.

A minority, but an increasing minority.

One vendor said that they were mostly buying CDR though.

Correct. But it happens to be an EDR vendor.

That's right. So that makes sense.

That makes sense.

I think this is intuitively sensible to me though, that an ounce of prevention is worth a pound of cure.

That's a very optimistic take. I think the pessimistic take is that the posture problems should be solved by now.

You would think.

The first CSPM vendor, if I recall correctly, probably was EvidentIO, I'm guessing, launched in 2013. We're talking about 11 years of CSPMing.

It's a long time.

Right. But it's a useful thing people buy today, 10 years after the market's birth, which is also—

Well, that's not too surprising. We still buy firewalls. We still buy AV. We just call it EDR now.

That particular take definitely offends me because I want to focus on the DR part of the EDR.

But— Listeners, I knew exactly how triggering Anton would find that one because I've done it to him before and it works.

We do buy that. We do buy firewalls. We do buy antivirus. We do buy other things. But cloud is supposed to be different and beautiful and good.

Well, what do we always say? So we beat on boats against the current, borne back ceaselessly into the past.

Okay. That's— we should probably like start climbing up from that low point. But you're right. That quote really clinches it. Yes.

Yes. Shout out to my high school English teachers for making sure I read my Great Gatsby. Thanks, guys.

What about the other buzzwords?

What else did we see?

ZT, ZT, zero trust is there, but ZT is not—

It's in the back seat.

It's in the back seat, yeah. ZT is not staring from every booth. I mean, sure, there's more attention in the government, but ZT came up a few times. There was one vendor with zero trust plus AI, because you want to use last year's buzzword and stick it together with this year's buzzword and then duct tape them together.

Oh, God.

That's there. What else is fun from the buzzwords? We did talk about XDR. We did talk about some of the SPMs that go— actually, even DSPM, the data security posture management, was quite visible. There were big booths of vendors that—

Several.

That's a problem that would come up probably even worse as we try to secure AI. Not that I'm trying to hint that you should ask me a question about my presentation, but the securing AI likely would be about securing data. Not largely, but like, A big part of securing AI would be about securing data.

Can we actually talk about some of the evolution we're seeing in the SIEM market and the chainsaws that are being taken to what used to be a coherent whole?

Yes. This was one of the things that counts as new, new, new idea. It's genuinely new. So, okay, but Tim, I can't really talk about this without mentioning my blog one more time.

That's fine. I'll allow it. One more time.

One more time. So there was this blog I wrote about decoupled SIEM. Which is basically separating the security part from the data platform.

Okay.

And it's kind of a debatable thing, and some people love it, some people think it's a bad idea because they want a tightly integrated thing that just works. But some of the vendors very clearly pointed out that they want to go hard on the security detection content tooling, maybe workflow, and not deal with the data platform at all. Basically, have customers do whatever they want. Whether it's BigQuery or something different.

And you think they can get away with that? They can actually build arbitrarily good stuff on top of underlying uncertainty?

And I think that my original take that for majority clients, it's a bad idea, I'm starting to waver. I'm starting to not go with that. And either my 5 years of leaving Gartner, my influence of the Gartner thinking and seeing clients has dissolved, or maybe things are genuinely changing, namely that majority of people wouldn't automatically go with an integrated. I think this is more TBD today rather than like, if you ask me that question, hey, do you want to know a SIEM where you need to install your own storage? I'd be like, no, that time is gone.

Right.

And when the early SIEMs had like, hey, install the database first. Like, the integrated essentially won.

Sure.

But now we are kind of in this next wave of, well, actually, maybe you want to deal with somebody who knows security really well and doesn't know data storage, and you would use— so I'm less adamant.

I think that's interesting. Now, you were so adamant 6 years ago that integrated SIEM was the way to go, you joined Chronicle.

Correct. That's exactly right. And I was very much convinced that for any kind of mainstream adoption, You have to have both platform and storage that magically support each other.

You voted with your feet on that.

I voted with my feet and in that sense with the wallet to point out that Chronicle offered a really magical combination of storage based on, of course, Google technical infrastructure and the security knowledge. So to me, that was a big win for the integrated.

Do you think we've lived up to that promise?

That's an interesting question that may go into philosophical debate for the next 2 hours, but the short answer is yes.

Yes. I think we have too.

I think we renamed— of course, this is also a big—

Oh, this is important. This is a good announcement. I'm proud to announce this one, listeners. We have renamed Google Chronicle to Google Security Operations, or Google SecOps for short.

Correct. And I think that it's very easy for me and Tim to take a very optimistic stance to it because we love the tool. And of course, Tim's also—

And of course, I've also voted with my feet and joined the SecOps team. Yes.

And that to me is still a very strong argument that if you want a modern SIEM, Most people would still go for a tool that harmoniously combines platform and security knowledge.

And intelligence, response.

But I can also start to see the scenarios where you can choose the specialist in security and commodity storage. It's sort of like, it's less offensive to me as a choice than it was 5 years ago, and that's for sure.

And is that because commodity storage has gotten better or because we see the incumbent players have become longer in the tooth? What's changed?

To be honest, I don't know. That's a really good question, which I do need to think about it. So if you—

I get a good question award today, listeners. I'm proud of that one.

Yeah, but the problem is that I was about to say I need to think about it and write a blog, but it would be a 5th mention of my blog, which I think is far outside of the anomaly.

We'll link to the blog in the show notes.

Far outside of the anomaly.

So what else was exciting this week? You had your first Waymo ride.

Oh, yes, that's right. I had my first Waymo ride, which—

What did you think?

Called with an app, like you're calling—

Just like an Uber.

A human. Yeah, human-driven. So of course, I read the announcements, and I watch the news, and I kind of know we did this. But it kind of still felt like it's a bit of a magical experience. It was actually a magical experience.

It was actually magical.

It was actually magical. And Tim was teaching me some tricks with, like, make the experience more optimal. Don't stare at it. Don't stare at the Waymo car. And a few other things. But it was actually quite magical. And some of the— it also felt like it's something that— many, many years of the work by the best of the best of the best people on the planet would produce the result.

It felt safe, right?

It felt safe, and it felt very like— well, magical is the word. It's interesting that I'm kind of sitting here gushing about it, and maybe people who had many Waymo rides are like, what is Anton talking about? But it does have the magic.

No, I've taken probably 20, 25 Waymo rides. I had one that was like a pretty bad failure mode, but It wasn't unsafe, it just, it couldn't figure out what to do and it basically lost its little mind. The poor thing. I felt really bad for it, but also then I was late to something and I was a little mad. But for the most part, it's amazing. It really, like, sometimes I think we get a little skeptical on the show about AI. I certainly fell out of my chair laughing last week talking about the risk of AI. But I think Waymo is a very impressive realization of what AI can do. That's productive. And so I would not be surprised to see Waymo-like experiences built into sim tools over the next 5 years.

Where you would be kind of a self-driving—

Yeah.

So I would believe in more elements— sorry, Ali— I would believe in more elements of self-driving SOC than before today.

Yeah. Having seen a self-driving car. Yes, correct.

This is really interesting. That to me is an interesting, like, a physical world fact from that land.

Yeah. Okay, so—

Presentations.

Presentations.

So this was the time when we didn't really, like, didn't have time for a single session.

No, I've been just doing customer stuff the whole time.

Yeah, unfortunately, I am in the same bucket except for 2 sessions.

Sessions that you gave?

Yeah, I know, right?

Anton, tell us about the sessions you gave, please.

I think at this point, I'm like, I think I'm being a little too self-serving. But I want to focus on one specific—

This isn't even a softball. This is like you tossed the ball up in the air and then hit it with your own bat.

Sort of, yeah.

Yeah, OK. That's fine. That's fine. Go on.

Go on. So we did talk with Taylor from our team, Office of the CISO. We did talk about—

Former guest.

Former guest on the podcast too, about securing AI. Specifically, we picked 5 fun risks.

5 fun risks.

Because everybody shows up and says, give me the top risks. And we tell them, well, top risks depend on use cases, kind of a little like SIEM. your top risks for your specific environment, your business, your use of AI, your models. And then this, like, I mean, the whole it depends answer is the honest answer, is the correct answer. It's also not very useful. So we came up with a list of 5 fun risks, which we think are exciting.

Fun risks?

Okay, we didn't call them fun risks.

We'll make them fun.

Yeah, okay. It's just basically us subjectively. Okay.

What are your 5 risks, fun or otherwise?

I don't think we have time for all 5 of them, but we definitely probably— Well, what's your favorite? My favorite is probably the risks with output.

Okay.

And these are kind of more conventional because people expect to see risky outputs or safety concerns or whatever other more classic risks. But the interesting bit about this is that unlike many other risks, which are solved by data governance, by other controls, by additional tooling, I think most of us have this suspicion that the output risks can never be 100.0% mitigated by anything. Well, you can mitigate some risks. You can—

Sure, but you can't get to 100% on that. That's the nature of these systems, right?

Correct. That's exactly right. And this is what makes it fun. This is what made it to the fun list. You can DLP, you can filter, you can do—

You can feed one LLM's output into another.

You can do all sorts of fun access controls on the inputs, outputs, conclusions.

Multishot prompting. Sure.

Incredible things. And they would make the output really, really safe. And would reduce leaks, reduce other risks with output. But it's probably a risk which is not really like— there's no fix-it button, whether easy button or hard button. There's just no button. You can contain it, but you can't control it. I think that's how we put it in the preso.

Interesting. This reminds me of a conversation I was having with Rob last night where we were talking about the step function change that you get in Chrome OS versus a more traditional OS.

Oh, yeah. Huge, huge step change. You almost have to— you probably, if you use Chromebooks for 5 years, you may have to read up in Wikipedia what's malware. You have to be like, what do you mean malware?

What is malware?

How can it be? The whole concept of malware would make no sense.

To a Chrome OS person.

To a Chromebook. To an entrenched Chromebook user for 5 years, you'd forget that thing existed. You'd be in the news and you'd be like, What do you mean malware? Like, like it's software that runs on that thing that you don't want to run.

How could it do that? How could you run software I don't want to run?

Exactly. Like it would be like incoherent concept to a Chromebook user. And this is not like an ad planted by Chrome team, is it?

No, it's not. It's not. It's not.

It's Rob and I were talking.

It is a fact. And we were talking about like architectural advantages and the durability that you get out of making good design decisions upfront, like how Chrome does software. And I think LLMs present a similar thing of we've made a design decision to accept the output of a system that we struggle to reason about and do not have a way of putting guardrails into. You have a different trade-off there versus a more traditional way of building software.

Yes. And I think that also brings up the— it's easier to convince people that a system, that a certain threat factor or threat scenario just doesn't exist. It's like, It's not even a no, it's NA. It's like, no, you can argue with, but NA, you can't. So with these risks, we have to deal with them in all sorts of exciting ways, but there's no fix it button. And of course, we talked about model theft. And I think the model theft was another fun one from the 5. Ultimately, the reason why we felt the model theft is fun is because models are a very portable store of value. Like, if you remember this story that I'm obscure that you told me about somebody who was concerned about their model being stolen while being exposed to clients.

Yes.

And the numbers that they mentioned—

The numbers were shocking.

The metaphor we used in the prezo, that if somebody steals iPhone plans, they're not going to become Apple. They cannot make iPhone in the garage. They cannot make iPhone anywhere. And even if they do, they still won't be Apple. It's not very stealable. But if you steal the model, you will have almost entire value of the company that built it at an inordinate cost.

Yeah, it's the difference between iPhones and the formula for paint.

Yes, that's a fair analogy. So to me, this was very fun. And that's why it made it to the fun risk. The other 3 I wouldn't mention, but the video of the presser would show up online at some point.

Excellent. Excellent. Well, maybe on that note of referring people not to your blog, but to your video, we can wrap up RSA 2024. And we look forward to seeing you on the next Cloud Security Podcast by Google.

And now we are at time. Thank you very much for listening. And of course, for subscribing. You can find this podcast at Google Podcasts, Apple Podcasts, Spotify, or wherever else you get your podcasts. Also, you can find us at our website, cloud.withgoogle.com/cloudsecurity/podcast. Please subscribe so that you don't miss episodes. You can follow us on Twitter, twitter.com/cloudsecpodcast. Your hosts are also on Twitter at Anton_Chuvaikin and _TimPico. Tweet at us, email us, argue with us, and if you like or hate what we hear, we can invite you to the next episode. See you on the next Cloud Security Podcast episode. Security.

Security.