EP245 From Consumer Chatbots to Enterprise Guardrails: Securing Real AI Adoption

Security.

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, UEBA, and Risk in Google SecOps, and of course, the one and only Anton Chuvakin, a reformed analyst and senior staff in Google Cloud's Office of the 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, what are we talking about today?

So I want to disrupt our flow and basically say, let's use this to do a call to action. And the call to action would be this. Based on this episode and many other episodes, our podcast should be renamed into AI Security and Detection and Response Security Podcast. Because honestly, I feel like we switch from securing AI to AI for security to SOC and SIEM, and then go back to securing AI, AI for security, SOC and SIEM. So at the risk of— and this episode will repeat the pattern, but maybe we should do a call to action to the audience and say, hey, Yeah. We are not at all tired of doing AI security, AI for security, and detection and response. These are lovely topics. We love them. We have fun with them. We deal with them every day.

Yes.

We should probably do something else too, right?

Are you asking our listeners if we should get back to our roots of doing cloud security?

Maybe people want to know how to use CSPM. Maybe people want to know how to patch Windows. I don't know. Like, call to action is the point, which is the, please tell us what should we cover? Don't give us advice to cover securing AI and detection and response because we'll do it anyway.

Yeah, that would be wasted advice because at this point you can certainly count on us to do that. But listeners, if you want to see us do more content about securing your containers or managing your identity lifecycle, do let us know because Anton is right to observe that we have perhaps strayed from our roots and we are off enjoying these new green pastures, but That's not to say we still don't care about cloud security. I mean, surely we've done 1 or 2 episodes about that this year, right, Anton?

Correct. No, no, we do deviate. And I mean, we did a great episode with a TikTok security lead that was mostly about compliance. So there are definitely exceptions. But the point is that they feel like exceptions. They don't feel like what we cover. And we seem to cover mostly DNR and securing AI.

DNR is perennially fun to talk about. What are we talking about today, though, now that we've had our—

Securing AI. Securing AI. Ah.

But you know what I think was so interesting about this episode is Thinking back on it, there are moments where it really sounds like our guest doesn't want to be in the business of securing AI. He wants to be in the business of enabling AI. And I think that distinction is really interesting.

Yes, correct. I feel like he made this excellent point about if there are 5 startups that secure AI, nobody sane would ever buy 5 tools to secure AI. That's just absolutely absurd. So there would be a platform that does more than that. look, it happened in the cloud. In the cloud, you just buy Wiz. Oh wait, we did. In any case, the point is, oh boy, you buy a single tool to cover most cloud security needs, and then you cover the gaps. With AI, maybe the same would happen.

But I think it goes beyond, you know, in cloud, sure, you buy Wiz, but Wiz doesn't make your organization, you know, better at using cloud. It makes your organization more secure at using cloud. And I think what was interesting about where our guest ends up going is he doesn't want to just do that. He wants to make you better at AI in addition to secure at AI. And I think that's a really interesting mindset shift for somebody who's spent his whole career doing security, not doing trusted secure enablement. So listeners, maybe with that theme and that call to action, get on LinkedIn, email us, argue with us, all of that. Let us know where you want to see things going. And with that, let's turn things over to today's guest. I'm delighted to introduce today's guest, former colleague of mine. We are joined today by Rick Katja, the CEO and co-founder of Witness AI. Rick, thank you so much for joining us today. I want to talk about the wave that we are all living in, which is the AI adoption wave. And you've been, not to call you old, but you've been around the tech industry longer than me. How is it different this time? Because I've seen, you know, this wave and that wave, but you've seen more waves than me. And so what's different this time?

Yeah. First off, thanks for having me. It's great to see you and Anton again. It's been a while. Things are going really well at Witness, so I'm happy to be here and talk about this. So, I would say a couple of things. One, the closest thing I've seen to this, you know, you gave me that little dig about being old. I'm very old. And my first startup was the first commercial web application server. So, this was, like, in the dot-com time. And I would tell you, in my opinion, enterprise AI adoption is probably closest to the web than anything else I've seen. It's different in some ways. And mostly because if you think back then, and I'll get to, I'll answer your question, but just kind of a bit of context. If you think back then, every company, you know, in the late '90s would say, I want to webify my company, web, web, web. And you'd say, what do you want to do exactly? Like, I don't know, web, web, web. And you get a little bit of that today with AI. You talk to the companies and AI, AI, we're gonna use AI, we're gonna use agents. What are you gonna do with these things? I don't know, but we need to use AI. So, I think it's interesting—

But also, wait, sorry for interrupting, but the speed is much higher though, right?

Much faster. That's it. The speed is much, much faster.

Yeah, that's also that plus what you said, plus hyperspeed.

That's right, plus hyperspeed. And I would say one thing that's interesting is if you roll back that many years to the web, And a lot of times the IT, the tech guys and the vendors helped the business people figure out what to do. Like they didn't really know. And so you'd have, like, I remember talking to CIOs and they would tell us we're driving the business because the business doesn't even know what can be done. They just know they want to do something. And you see a little bit about that here. So what's different? So I would say a couple of things. First off, the most obvious one is every single CISO I speak with says, I wanna be Dr. Yes, I don't wanna be Dr. No. I am pushing this. And if you think about something like cloud, it seemed to me cloud always seemed like this battle between users that wanted to use cloud and security that wanted to squash it everywhere they could. You know, the CASB were mostly about blocking stuff, I think, more than anything else when they first came up.

We'll get to CASB in a minute.

Yeah, CASB would make it a pain in the ass.

Okay, I'm sure we will end up. But I think one thing that's different here is, I've never in my whole career, and worked with CISOs for decades, I've never seen a scenario where they are aggressively saying, we need to help the business and help the users do this. So that thing is one that's very different. I think the companies are very unsure about how this is going to go and what they want. From a vendor perspective, so I was at Black Hat 2 weeks ago. And we met with, I don't know, 2 dozen CISOs, and they'd say, agentic, agentic. And say, oh, what do you want to do with agents? And they go, I don't know, but we know we need to have agents. In fact, there was only one company we spoke with where the CISO could speak with any depth and clarity of what they were really going to do with agents. Everyone else sort of knew this was coming, they had to do it, but they really didn't know. The next thing is what we're finding in practice, this part's really different, is I would say conversations are not queries, and models are not databases, right? I think, you know, if you talk to these LLMs, they're wrong, what, 15%, 20% of the time?

It depends who you ask, but it probably is in that bucket, right?

In that bucket. I think if a SQL database gave you wrong answers 20% of the time, people would go crazy. It would be worthless.

But yet, Can you imagine?

Can you imagine? Like, hey, what's the inventory on these cars? I don't know, 100? 50?

Nah, it's 80.

1,000?

A billion?

But people are deploying these things anyway. So, you know, conversations are unstructured, whereas a query is often more structured. The thing answering on the other end goes in weird directions. You don't know what's coming back. And so what we're finding is, and, you know, in the same way that The pattern matching you used to use to kind of protect data and control things doesn't work very well here. So, despite people not really knowing what they're going to use it for, having a high error rate, unstructured, surprising responses, CISOs want this stuff and companies want this stuff to go. And so, it's all going into production anyway. And in some ways, you have quite a bit of chaos. And we see that with our deals. It's a good chaos, but it's a chaos. So, I think that part is different from what I've seen before, is they don't really know what to do, but they really want it. The teams that would typically, their function was to sort of slow this down and block it, are trying to make it happen fast. And it's unstructured and surprising and unexpected in ways we haven't seen before. And then a bunch of things fall out of that. So, I'll stop here, but say a bunch of things fall out of that, problems that we're actually seeing.

I think that's such an interesting observation about being wrong, because I've said the same thing. Like, imagine other sort of labour revolutions we've seen from technology. What if CAD had gotten it wrong 20% of the time and, oops, some of our bridges fall over now worse than they did before because, well, we saved a lot of labour though with the CAD. Or if QuickBooks didn't tell you how much money was in your bank account, or if your spinning loom made pants instead of shirts. There are so many things here where technology has improved accuracy. This is one of the first times I can think of in history where technology's gone forward while reducing accuracy.

I think this is like, this is gonna derail the planned show, but wasn't there like Crossing the Chasm book basically said that the innovator shows up being better in some area and likely worse with others? So I think that's actually not unheard of because of that. Like I remember to use an example from Rick's and my childhood, Namely, SIEM vendors that based their technology on a relational database being clobbered by the search indexers. Search-based people showed up and said, here, we have that. And everybody's like, haha, you're idiots. You don't even have a schema. And then many, many years later, things changed, to not go through 20 years of SIEM history. But being bad in something like—

My ears are doing a lot of work, Anton.

Yeah, yeah, yeah. That's a longer story. But I mean, being bad in something like accuracy is probably okay because there's some advantages OK, I'm not going to defend LLMs. Let the robots do it. Let me get to the question.

Anton, don't worry. The robots will hear this. They'll remember it. They'll take care of you when they come for power.

Oh, lovely. Thank you. Thank you, robots. So in any case, one thing that came up in this whole adoption wave that Rick just described is that employees bring this tech to work. And we've covered extensively shadow AI, like shadow IT but with AI. Right now, I'm doing a research piece on shadow agents, what will happen if employees bring agents to work. Which would be really fun. But what is your take on this whole consumer-grade AI versus business AI? Can you magically turn one into another? Like, what's your take given where you are at Witness AI?

Okay, so with Witness, we're selling primarily into big enterprises. Like, you know, right now I'm at like a top 3 airline, top 3 automaker, top 3 retail financial, top 3 utility, top 3 telcos, so big companies. And ones that you would expect, they're all regulated industries. You would expect they would move slowly and they would have enterprise, not consumer-grade technology. But here's what we're seeing in practice. There's sort of 2 sides to it. So the consumer side of AI, the way that's affecting the enterprises we talk with is probably the way you'd expect, which is every employee hears about some cool new toy, some new app.

Mm-hmm.

And they use it. And most of these orgs, the more reactionary security departments may say, that's no problem, we'll just block it all at the firewall.

Yeah, which of course forgets that their mobile phones exist.

Like, how are you gonna— but even worse than that, like, how do you even know about this to block it at the firewall? They don't know.

They block all on the firewall.

Yeah, everything. I've only talked to one or two companies in parts of the world while I've been here witnessed that could credibly say, don't worry, our employees will never do that. Every other place, they're doing it. So the consumer side is just like maybe with cloud, employees are using consumer tools all day long. And so you have some issues around IP loss, data loss, also what those apps might be saying back. So there's the consumer part. is touching the employees because they're using consumer tools. There's a second part, which is the companies themselves are standing up their own models. And I tell you, so far that looks a lot like enterprise technology. Like it's going slowly. I'll give you an example. We met with the CISO of one of the biggest investment banks in the world, like end of 2024, Christmas. She came out and we did this private thing with a bunch of other portfolio companies. And we said, how many AI projects does the bank have? How many projects are going? They said, oh, 500 AI projects here at the bank. And we said, how many are in beta right now? And they said, 13. We said, by the end of next year, how many will be in production? And they said, zero.

Zero.

Zero.

Yeah, it'll easily be zero.

So, I would say the enterprise AI side, and this is a big, rich, fast-moving bank, is moving at enterprise pace. And that's a different issue. Where they're seeing more issues is where the employees are touching the consumer services. I mean, we're literally tracking 5,000 of these at Witness that we have controls over. And I know we're scratching the surface. Like, there's probably at least 2 times that that we're catching up.

Are you basically saying that the battle for shadow AI governance is lost? We should just go home? and get a drink. Like, that's what I'm hearing. Like, there's no— the speed of official adoption would always be so slow compared to ad hoc shadow AI style adoption that it's unbeatable. Like, that's what I'm hearing.

I don't think— well, so if I thought it was unbeatable, I wouldn't be running Witness. I think it's beatable. I think it's a tough problem. I think you have a long tail problem. So, you know, if you look at the top 1,000 apps we see inside of enterprises, The bulk of them are 10 apps, right? It's the ones you'd expect. It's ChatGPT, it's Gemini, it's, you know, you know what they are. There's some unexpected ones they don't quite think about. Like, I don't know that I want people pasting things to Grammarly, but generally it's the ones you know about. Then there's some long tails. Then there are ones that pop up. Like there was a whole DeepSeek, you know, panic for a while.

Yep.

And so you can get a good handle on the shadow AI on the consumer side. And it's not that hard to get a handle on the shadow AI on the enterprise side because you're Developers probably aren't standing up 1,000 models, they stand up a handful. The place where it's about to get bad is the agentic side, like you mentioned, because the next place where it's going to explode is people standing up agents that carry their credentials and authorizations and can touch file systems and their cloud systems and all that stuff. So basically, the way we see it is, At the high level, you have, the way I would call it, you have kind of a confidence problem for enterprise AI. Can we confidently deploy this stuff so my employees can use consumer services, so my employees and customers can use enterprise services, and God knows what's happening with this agentic stuff. But so you have this AI activity problem with employees touching the shadow AI part. How do I secure, give them sort of training wheels or— Those bowling alley bumpers for consumer stuff? How do I have security when I stand up my own models? And then how do I control workflow and process when people stand up all these random agents? And let me give you one example on the security part. It's pretty funny. So we have a bunch of customers who are standing up their own B2C consumer-facing chatbots.

Yeah, of course, like support, likely.

You know, they all have these things. The hotels are standing up bots. Everyone's standing up something to do tickets and planning. I would say almost every single one we've spoken with, do you know what the first thing that happens when they turn the bot on in testing is?

No.

Now I'm super curious.

Recommends a competitor.

Wait, wait, a competitor for their— not for the— so basically you ask a support bot for car company A and it says consider a car from car company B because they're great? Or like, I can't imagine.

Something like, hey, you know, here's my budget, here are my needs, I have a family of 4. I'm looking for something that's safe and effective. What's the best automobile? And you ask the bot on Car A, and the bot says, well, Car C has a great car that meets your needs. Or you go, hey, I'm going to go to Puerto Vallarta. I'm going to go to Puerto Rico, and I'm looking for a hotel that has this, that, and the other. And Hotel Chain X's bot says, well, Hotel Chain Y has great deals right now. And what happens is, you know, these models don't have identity.

It's not your and my problem. It's the business guy's problem, right?

But, well, if you start saying, look, can I deploy this stuff with confidence? Like, so we look at it at Witness, we say, look, this is not a FUD security guy in hoodies. I mean, listen, you and I were at Chronicle, like, we know the whole don't use the guy in hoodies, like, don't do the scary security marketing.

Right.

And I think with a lot of this stuff, if you take a step back and go, it's not so much a scary security problem, It's an adopt things safely and with confidence problem. Then, you know, my B2C chatbot recommending my competitor because it doesn't know that it's Company A's chatbot, because the model doesn't know that, and they didn't, you know, that's a part of adopting it safely. Having my employees with best intention upload a bunch of confidential information to Gemini or to ChatGPT and say, hey, help me format this into a great letter for the customer, or, hey, I'm a call center in another country and help me with my English, help me make this response look good, but there's a bunch of confidential data in there. Like, the best intention, but you may have a data loss issue. Like, these are all safety confidence things and not— you're right, it's not security per se, but the CISO sure has to worry about this stuff.

Okay, so this is kind of melting me because until a second ago, I was gonna say, Rick, you sound like Kasbian about 2014. Is it different this time? But it actually sounds like it is different this time.

It is different.

But it's also not decouplable from— it's like my dream is to decouple this stuff from proper security, privacy, compliance stuff, but it's almost impossible. And when we present to clients, we say kind of the same thing, that broad domain, you can't always tear irrelevant things apart. And that's a confirmation that ultimately the cyber guy would be stuck with some of this.

Look, I'll give you one, one thing about how we thought about the problem. So when I went to do the Series A, I talked to some of the, one of the VCs and they had one of these AI security market maps and it had 18 different segments.

Oh yeah.

And there was like 10 startups for observability, 10 startups for policy, 10 startups for detection, 10 startups for AI DLP. And look at this, like, you guys have this wrong. Like, this is a safe adoption problem. If you look at it the way you guys are looking at it, you're saying a CISO needs to buy startups from 5 products so their employees can use Gemini safely. Like, no one's going to do that.

No way.

No way. It's crazy. But if you look at this as a safe adoption or confident adoption of AI in the enterprise problem, It's much broader. The feature set is different. It is kind of brain melting, Tim. Like, it's not, let me just block stuff.

So this kind of gets to the second half of what I wanted to ask you though, which is all the stuff you're describing. Couldn't I do this as a SASE or as a secure browser? Like, how are you going to have this be something that is both budget and attention keeping in the long run? How's this not itself get rolled up? You're making a roll-up argument. What's the next roll-up argument counterargument?

Yeah. So our take here was twofold. One, I do think this becomes, every time there's a big major shift, you know, a bunch of companies get rolled up and one company kind of makes it to the other side.

Yeah.

You know, Okta on identity, CrowdStrike in EDR, Netskope in CASB. I was going to say Wiz in cloud security, but maybe not, maybe not, not anymore. And I do think one company ends up being the next big platform. We think the opportunity is here. I think one of the reasons why the existing things don't work well, and we plug into them and work with them, they just don't understand how different this is in the way the conversations flow. They view the world as a much more structured, strict thing.

Mm-hmm.

They may say, well, this data is confidential. I use DSPM to tag my data. And the notion that I've cut it out and pasted it, I lose my tags, or that stuff I can't apply patterns to, they just don't work that well.

All patterns apply in English, but AI would work just as well in Hungarian. And that came up with some kind of naive soul who's trying to filter the input for the chatbot without— forget multimodal. They even forgot the bot can understand, maybe not any language, but can understand several languages, including those for which rules don't exist. Like, forget the screenshots and forget the images. Even the prompt in another language confuses a good bunch of classic tools applied to AI, right?

Yeah, at least in the era of filtering for SQL statements, we only had to filter for SQL. We didn't have to look for SQL that was written in French.

Guys, it's worse than that. The latest thing is the emoji attacks. Like, you can basically write prompts using emojis that the model will understand, and it will cough up data Based on a combination—

Wait a second. The emojis—

That's amazing.

I'm going to read up on this. If you have a link, send me, because I thought that the video example or screenshot example is as far as we go confusing the filters. But emojis sound even more fun.

Yeah, tell me how you're going to regex emoji prompts. It's just not going to happen, right?

I've got a funny story about emojis. I once was responsible for an OMG at a top 3 American bank using a single emoji. My product was responsible for onboarding users to a major financial product. And one of our customers signed up as Happy Plant Store, plant emoji. We passed that name verbatim back to our financial ecosystem partner, and their mainframe was not prepared for plant emoji, and they fell over and had like hours of downtime while they traced plant emoji.

Emoji, the famous emoji injection attack.

Exactly. Exactly.

But, but so like literally you see people doing this.

So same problem.

Same kind of problem. Like, we, you know, our take is you have to build a system using AI to understand what people are doing with AI and to detect it. Like, our stuff does this inline, on the fly. And it was really hard to make it work.

You know what makes me nervous though, right? I'm nervous because it sounds like filtering, right? Like, it sounds like you would filter your way to success here. And I'm nervous when people say that for the reasons we just discussed, right?

Some of it is filtering. The most common use case we see, like, here's where all the Like the CASBs with DLP fall over. You put something in a prompt and it's got a bunch of, I don't know, customer information, and the DLP goes, ah, block. Okay, block.

Mm-hmm.

But remember, we started this talk by saying that CISOs want to be Dr. Yes, not Dr. No. What they want is they want you to, on the fly, figure out if that thing is redactable, redact the pieces, Send those things out with tokens that can be processed to the LLM, get an answer, come back, unredact it, and give the user the answer. Like, we do that. I haven't seen anybody else do that. Or another example is to say, well, I want my users to use Cursor or Visual Studio Code and GitHub Copilot. I want my developers to use this stuff, but I want the answers to come from an internal LLM trained on their code. There, I want to pluck the prompt, you know, off the wire, scan it, and reroute it to a different model, even though I don't own the endpoint app and I don't own the model. Those kind of things— but your CASB can't do that, and we do that. And so it's not so much filtering and blocking, though that's an option. It's how do I give people a way to get answers, do their work, get stuff done, do it safely, even when they don't really know what they're doing, And tell the security and privacy org, don't worry, there's a little silent angel we've put on everyone's shoulder that's making sure they don't accidentally get you guys into trouble. That's pretty fun. I mean, the technical problems are pretty interesting. To make it work, you basically had to build a super high-scale network services platform, a bunch of super high-efficacy AI engines that run inline, and then—

Which cannot be tricked by the same tricks, hopefully, right? Because otherwise, it's the classic Your firewall that stops buffer overflows have an overflow, right? Your engines should not have the same problems in LLMs. And I don't know how you would achieve that though, that your LLM filter for LLMs doesn't have LLM systemic issues.

Well, so a lot of that's tuning and design. We've done that so that you have small models that run well. The other thing is, let's go back to the 3 things I talked about. So there's the employees accessing consumer services. Employees and customers accessing enterprise services, and then agentic. Our notion is you want one backend that can do observability, policy control, and protection across all 3. And if you get that right, and none of the existing, you know, if we go back to why can't the CASBs do it, none of them have been deployed in all 3, like one for internal only, one for external only, and then one for sort of crossing the network boundary. So, building something that can do all 3, we think, is the win. We think those 3 together is gonna give the opportunity for a platform. I mean, the business growth seems to support that, but that's our take.

So, Rick, there's—

Is there an—

Oh, go ahead, Anton. Go ahead.

I was just curious about the identity angle because it comes up in many agentic conversations because people assume, I mean, even the, is agent a user or is agent a workload? I've managed to cause a minor Twitter war and then minor LinkedIn war over that topic. Because it turns out it's contentious and one half is completely convinced their answer is the only possible answer. And the other half is absolutely convinced that their answer is the only possible answer. And it's inconceivable that somebody would argue the opposite point. And then the vote was like 40-60 or something. So, how are you approaching the identity side of this? Because it sounds like a fun one in addition to the stuff we discussed.

Yeah. Yeah. So, there's 2 parts on the identity. Well, actually 3 parts. There's the easy part, which is the employees using commercial services, consumer services, which is, you know, you'd think, we didn't think this was that breakthrough of a decision, but like we made the decision to say, well, we only work by connecting to your directory services and sucking in your existing identities and groups. And we got a lot of pushback from that, which is really weird.

Yeah.

But the nice thing is then you can say, we reuse the identities that you have so you can build policies based on the groups you have. But that's pretty straightforward. There's a second identity, which gets to that recommend the competitor, which is, how do I apply and enforce an identity on the model itself? Like, no, no, you're hotel chain X, you're not hotel chain Y. And there's some interesting things we do kind of behind the scenes with system prompts to enforce identity on the model. But I think what you're asking is the third one, which is, Those two aren't that hard. It's the agentic identity really blows a lot up because—

Yes, that's the fun part.

A person has a lot of authorizations that you wanna be very careful that those just sort of blanket get applied to some agent. And, you know, as my co-founder, Gil, our CTO, says, he's like, hey, you know, I build an agent, it's got access to my file system, it's got my cloud-based credentials, You know, I go up and get a sandwich, this thing can wipe my file system, it can kill all my, kill everything in the cloud. So I think people are really starting to try and sort that out of what should an identity for an agent be versus the person it may be representing. Our own view is you need both. You need to be able to have an enforced system where the agent picks up the user's authorizations. You also need one where they have a subset. We also believe, and the way we're building ours is for certain things you want human approval in the middle. And the way this is going to go is you're going to say for certain steps, like this can't get completed unless a person comes in and goes, no, no, final button. I pushed the button.

Yeah.

This notion that every agent's going to pick up like their developer's authorizations and just run around and go do stuff, it just seems not like it's ever going to work for real.

So I hate to do it. And we had so many other things we wanted to talk about today, so we'll have to have you back in the future, but we have to close. So do you have one tip to help people, you know, secure their AI? And 2, do you have recommended reading? And the reading can't be Anton's blog.

Oh man, can the tip be talk to Anton?

Yeah, that could be the tip. That's fair.

So I would say, here's a tip I would say is the basic thing. Look, we've deployed a bunch of customers where things go wonky is when they're trying to do too many things upfront. And I would say the first tip is, first, get observability under control. See what people are doing, because 100% of the places we've deployed, they are very surprised by what their users are doing, always. And before you go in, they're like, no, we know what they're doing, blah, blah, blah, they're not there. And then they put it in, like, oh my God, ah, it's not at all what I thought.

Exactly, yeah.

So, the first tip is, Start with observability. Once you have that, a lot of things become easier. The resource one, I didn't come prepped for this, so I'll give you— I don't have, like, a great blog. I can't give Anton's blog, so I'll go to something else. I was asked recently a book I would recommend if you're in any sort of growing company or building a company. One of my favorite books ever is The Soul of a New Machine. Have you guys read The Soul of a New Machine? Have you heard of it?

No. No.

No, I haven't. Tracy Kidder, he won the Pulitzer Prize for writing it in, like, the '80s. And it's basically a book about a company called Data General that isn't here anymore. And it's basically, they embarked on a crash program to build a new minicomputer to fight DEC at the time.

Okay.

Which also isn't here anymore. It's an amazing thing about— I like it from a startup perspective, how you build a technical organization, how you treat them, you know, like this notion of mushroom management, like put them in the dark, feed them crap, watch them grow. Like, where all these things came from, how you structure it so they get meaning. How you make decisions of do you hire people right out of college because they don't know anything and therefore don't know limits versus experienced people who can do it. And it's one of the most amazing books about building a company. I think of it as building a startup, building a technology, and how you organize for success. It's one of the best books I've ever read. And I would say if you're going to go start a big project or start a company, go read The Soul of a New Machine first.

Amazing. Rick, thank you so much for joining us today.

Thanks, guys. Great to see you. Thanks for having me.

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/podcasts. 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. Security.