SPEAKER_02 0:00

Hey everyone, welcome back to the Candid CISO podcast. I'm your host, John Donovan. Here's the thing keeping a lot of security leaders up at night right now. Your organization is standing up AI agents that can read data, take actions, even make decisions faster than any human ever could. And most of us are still trying to figure out how to govern that. Who's accountable when an agent acts? How you audit it? How you even know with confidence which agent you're actually talking to. Today's guest has spent his career on exactly this problem: identity at scale under real-world pressure. Boyan Simik is co-founder and CEO of Hyper, and he's been deep in the FIDO alliance, shaping how the industry thinks about authentication and identity management. We get into what it actually looks like to bring identity controls to AI agents, including a live look at attaching an evaluator agent to a red team session, and what that means for how you audit and govern this stuff in practice, not just in theory. If you're a CISO trying to get ahead of agenciai before it gets ahead of you, this one is for you.

SPEAKER_00 1:07

Let's get into it. Today's episode is brought to you by Hyper, the identity assurance company. As AI agents join the enterprise workforce, one question comes first. Can you verify who or what is acting on your behalf? Hyper helps organizations answer that question. Learn more today at hyper.com. That's HYPR.com.

SPEAKER_02 1:27

You know, you've spent your career on one question. How do you know who's really on the other end? Where did that really start for you? I mean, I'm sure it was before agentic became a thing.

SPEAKER_01 1:37

So Yeah, I mean, for for me, it was I worked in I worked in security, started my career as a penetration tester in the space, and then moved to running a security team at a financial institution. And most of the stuff that I dealt with on a day-to-day basis was something stupid happening with a credential. And it was just like, okay, there has to be a better way. And this was back when, you know, I remember my first day on the job at that place. Uh, I walked in and there were one of those quads um with a table in the middle where people could just kind of turn around in their chairs and have a conversation at the table. But in the middle of the table was taped a uh old school RSA TOTP token. And so the four people, the four people that uh worked in that cube, uh, everyone when whenever they needed the OTP code to do something, they would just push back off their desk in their rolling chair, they would look back at the token on the table, they would go back and type in the code. And I was like, oh man, I have my work cut out for me here. So um starting with that and and you know just the pain of dealing with lost or stolen credentials for years really started my really started making myself think, and and ultimately my co-founders here around like there has to be a better way to improve the user experience and improve security for the end users. And and that's really how it all started out.

SPEAKER_02 3:12

So you know you got me digging around in my desk store to see if I can find one of those old school RSA tokens. I have one around still. Well, there's a UB key still in the in the wrapper. I mean, this is one of those things where it's like we've gone to multi-factor, you know, we've gone to all these other things, but you know, the human in the loop is still kind of both a beautiful thing and a terrifying thing. How did you get on this whole thing with hyper and passwordless? Uh, you know, how did that transition really happen for you?

SPEAKER_01 3:42

Yeah, well, when we first started out, we we initially set out to set a capabil to create a set of capabilities to um eliminate fraud and and financial payments. And uh so we built an entire anti-payment fraud platform um that payments could flow through. And it didn't work out for multiple reasons. But one of the things that we uh we got a lot of interest from uh enterprises was the authentication system that we built into the platform. And uh back then it wasn't called passwordless, it was just called um uh non-shareable credentials for us because we wanted to make sure like what we wanted to do was create an environment where there were no shared secrets. People can't share a credential either intentionally or unintentionally. And so we created that system. Um, I met my co-founders here in New York City, went down that path, and then um, you know, really bootstrapped the company in our early days and and managed to get some early enterprise customers on the platform, um, companies that were like really on the bleeding edge. And this is back 10, 11, 12 years ago now, where um things like the FIDO standards were brand new, right? And and you were holding up a U.S.

SPEAKER_02 5:02

I want to talk with you about that, definitely, because you you have some experience with the whole FIDO alliance. So, you know, tell tell us a little bit more about that. How did how did you get involved there and and kind of what was the impact you know around that? I mean, again, I love this idea. We're talking about most people were still just trying to figure out how do you get a second factor in, right? Whether it's that UV key or you know, time-based, uh, you know, the the those RSA uh tokens were the main thing for so long, right?

SPEAKER_01 5:28

Yeah, like our mission was how can we build an authentication system where it has zero shareable credentials to it? And most importantly, how do you make sure that the human can never fall back to a shared secret? Uh and so uh what we wanted to do was create a system that did that. And in research of creating such a system, we actually came across the early drafts of the FIDO standards and the FIDO specs. So we implemented it like in the very early days. And this was back, you know, back then, um, there were like multiple open standards kind of competing to get market share. Uh, there was like this thing called the biometric open standard, there was FIDO, there were others. Uh FIDO ended up being the one that most organizations adopted and ultimately won that um battle. But for us, it was just very, very important that people should never have to type anything in and that they should never have to be able to share a credential. Uh and and today, if you know you show your YubiKey there, like that's part of the appeal of technology such as that, which is we can give this credential to our end user, and we don't have to worry about them like sharing it with somebody else.

SPEAKER_02 6:44

That's a pretty amazing thing. So you started out really kind of, hey, let's try to implement this phyto thing, and then you end up being on the board. Kind of how how tell us a little bit more about that. Because I've had a couple cases where you end up on a nonprofit board, and it seems really cool to a lot of folks, but you know, the reality is open standards are a labor of love.

SPEAKER_01 7:05

Uh, you really like you really have to uh you really have to be in it for the right reasons. Uh, and and ultimately I think the people running the FIDOLIN and focusing on on that work uh don't do it for financial reasons largely. They do it to improve the industry, right? They're the we're we're all combined, uh we're all a group of people who have all had family members and loved ones scammed in some sort of way by a credential, and like we just don't want that to keep happening. And at the same time, we know that identity is such a critical component of every single technological ecosystem that nothing meaningful tends to happen unless there is an open standard behind it where people can have full transparency into up how it operates and why it operates that way to ultimately adopt it and drive it forward. So I joined the board of the Fight Alliance a few years after our initial involvement there, had been uh on the board since then. And my job there and my role there is primarily making sure that you know I do what I can to keep us on mission. Uh, and and and that mission is eliminating passwords and shareable credentials.

SPEAKER_02 8:16

I love that. And and we are gonna shift a little bit here to our current agentic you know uh environment. But before we get there, I'm gonna put a plug in just for, like you said, labor of love. I know someone else who's uh a startup founder who put together this site called dontscamgrandma.com. And it's kind of all about how to test this out for your your your grandma. There's a few other sites showing up like that. Um so I guess the thing is when did you first kind of notice that that the that question, kind of what we're looking at, the subject of the question of who's on the other end was no longer a person? Like when did that start to become something that you know you were really kind of honing in on?

SPEAKER_01 8:56

Well, back in the early days, earlier days, they really just started with shared accounts and and you know, where the person on the other end is a person, but you don't know which person. And then it started moving to service accounts uh and and ultimately um uh you know uh like workload accounts. Um and then a couple of years ago, we started to see the first twilight of what I call non-deterministic non-deterministic systems using um using credentials, which is particularly concerning because if I have a Python script and that Python script has access to a credential and I run it off of some service account, um I can run that Python script a million times and I'm gonna get the same result every time. But if I if I give a credential to an LLM and uh give it, ask it to do something a thousand times, it's not gonna do the same thing every single time. And that's where this concept of non-determinism really came into play with regards to how we use credentials and the importance of securing that. And ultimately, uh as we here at Hyper started leveraging agents in a much more meaningful way. Uh, and by meaningful, I mean not just asking it to write some basic code or summarize emails for us. Um, we discovered that we needed a lot more confidence and control around how these systems are operating within our own environment in order to have confidence using them meaningfully, because we mostly service large global enterprises, and you know, we have a responsibility to stay secure.

SPEAKER_02 10:46

And we're gonna get more into that kind of you know, what happens when you give agents credentials and maybe the incentives. I mean, it at the time of this recording, right, the hugging face thing is the big thing that's in the news, and you know, the different uh frontier labs are are setting these agents out that are supposedly escaping and hacking each other, and you know, we can go down the whole kind of you know, we're we're all gonna die, you know, uh kind of scenarios. But before we get into that, I mean I really like the path that you're talking about, which is you guys were looking at how do you use agents yourself inside your company, and then because that's kind of the expectation. So I guess you know, in those conversations you've had with those customers, especially in some of these bigger organizations, you know, let's go through one of those scenarios of you know that that Monday morning meeting with the CEO. CEO just got back from golfing with some of his buddies, and they're all saying, Hey, we're doing all this kind of stuff. And so he walks in, and the CIO and the CISO are there, and you know, like we need to deploy agents. Walk us through kind of some of what you've heard, you know, change the names. We're as candid as we can be here on the candid uh pod uh CISO podcast, but we know you know there's confidentiality and and uh clients you need to protect. But it seems like this is a theme that's coming up, right, Boyan?

SPEAKER_01 12:04

Absolutely. It's it's really interesting how uh how many CISOs and CIOs even I've spoken to that just say, like, hey, like I just got a phone call from a CEO, and like we need to move faster, faster, faster. And all their peer, you know, my CEO thinks all their peers are going to outpace us and and uh and and we're no longer going to be competitive in the market as a business. So though it's interesting how the priorities can be different sometimes because the CEO wants to get the business value out of using AI everywhere, and the CISO wants to make sure that their company doesn't end up in the news. Um and and there's a real push-pull there. But I do think that the technology that wins out in this space is going to be the one that best enables companies to do both. Uh and I think that um as as we start impl as we start implementing agente controls, there are a couple of key use cases that I've seen firsthand that are just most critical. One is um, how do we make sure that we can maintain the same accountability model with our agents? Which is okay, we're gonna deploy agents as a company, we're gonna give everybody AI, it's gonna be terrific. But if that AI makes a mistake or does something it's not supposed to, what accountability model do we follow? And most companies right now want to maintain the same accountability model that they already have, which is um if Bob's agent does something stupid, it's Bob's fault. Um, we're going to we're going to hold him accountable to that. Um and that's the first thing that companies need to figure out, which is how can we have that breadcrumb trail for every single agent that's uh deployed in our environment? How can we associate a human identity to it and make sure that um if and when things do go wrong, we have the ability to talk to the right person and maintain that accountability model. So that's one thing. Um and then there's another set of capabilities that companies have very much on the for the top of their mind, which is how the heck do we uh restrict the things that agents can do on behalf of humans, right? Because we took we we talked about previously how you know humans are typically only aware of about 4% of their permissions and agents know 100% of the permissions.

SPEAKER_02 14:32

Yeah, we're gonna dive into that one a bit more. I do want to stay here just for a minute. And and you know, when we did a little bit of our our prep call, we were talking last week, you were actually at a banking conference, right? And you know, the banks, in my experience, and I've had the opportunity to work with a number of banks and financial institutions over my career, they both get very advanced with some technology, you know, they implement things earlier on, and then they're very conservative, which we hope you know is the case because we want to make sure that the money stays in the banks, right? Um what what was some of the buzz you got out of that conference? Like, what are those leaders actually afraid of? What are they excited about?

SPEAKER_01 15:12

Yeah, banks have it pretty tough because their business is how can we convince people to switch to our bank as much as possible? And banks traditionally have very high switching costs. If I was to switch my bank today, my primary bank to another bank, I would have to block out like two weeks on my calendar to do that. Like there's so much stuff tied to my bank account, there's so much legacy there. Um and and I think one of the things that AI banks are particularly concerned about is how can AI make switching costs of banks really, really easy. If I can switch from bank A to bank B in less than an hour by using a bunch of agents that just do all that stuff for me, that becomes very concerning for banks because they spend a lot of time making it particularly difficult for users to move off of banks. Um, so that's a big red alert for a lot of banking institutions. And then the second one is on the fraud side, which is if my user clicks some link and that link opens up their Chad GPT, and then their Chad GPT kicks off a workflow that transfers money out of their bank and sends somebody else. What is the accountability model there? Who is held responsible? Like, how do we prevent scams and things like that? Um, because humans will want agents managing their bank accounts. Uh, but to what extent and how can you limit that is very concerning for banks because they're very much like, hey, if this starts happening, what kind of consumer protection do we have to provide to billions of users? Uh if if agents are running their bank accounts.

SPEAKER_02 16:56

Yeah, it has these implications for kind of that whole contract. And you know, what I was thinking about as you were saying that is I haven't switched banks in a while. I've added banks for other things, but I did switch phones from one provider to another, and oh my God, do they also try to make that as hard as possible as well? So, you know, I think there's this second order of disruptions that are coming. And you know, again, what are the things that you know you're bringing up some really great things about accountability, whether it's inside your company with employees, you know, it it and and there's something that maps there well that you know, hey, the accounting department's still the accounting department, and as the CIOs and CISOs need to help to put the right kind of guardrails and gates around there to help people, you know, innovate, but not uh, you know, accidentally transfer you know stuff out as opposed to our old kind of you know phishing scams where someone would say, you know, send the money out. What happens when an agent, a very helpful agent, wants to go do that for you to you know, can't do that bank transaction? You know, I'm hearing that there as well. Um I think it's interesting, and and maybe there's a difference between the smaller companies and startups where it's still a YOLO environment, you know, one where you know I was running a security program, you know, I had all these people trying to use all these different tools. And so it's nice when the CEO goes and says, Thou shalt use whichever, you know, chat GPT, cloud, whatever it is. Now you have maybe one attack surface to focus on and say that you've got a standard set. I guess my question to you around some of this is how the heck do we you know put some of these guardrails in place? And it really feels like you know, yesterday's controls are not gonna hold up in this case where you know these agents are moving at machine speed.

SPEAKER_01 18:40

Yeah, it's uh it's one of those things where the controls that we put in place have to be living breathing things. And by living breathing things, I mean they have to be agentic in nature because there's no way that we can apply deterministic controls to a probabilistic system. And agents are probabilistic, and most of the controls that we put in place right now as businesses are deterministic in nature. Put in this firewall rule, block this port, like that type of stuff. That's not gonna fly. If you create a policy for an agent that says you can't send more than $5,000 in a transaction, it's gonna say, okay, I'm gonna send five thousand or six thousand one dollar transactions and just do it that way. So um we have to have uh probabilistic systems implementing uh uh implemented across a set of probabilistic tools, which is super scary for us security people, right? Because we're used to going and saying we have this much coverage across this many risks. And when that could change on a dime, it's a very different um threat model. So I think as security practitioners, like we really have to think about how do we change the way that we approach building security programs and start thinking uh in terms of uh the agility of the controls that we implement rather than necessarily what controls we we we apply to a specific um environment. And so how quickly can we train, how can how quickly can we adjust and and switch out our controls is is critical. In the cryptography space, we're talking about like crypto agility. Like we need control agility for our AIs. We have to move very, very quickly uh to modify them.

SPEAKER_02 20:31

I want to get back to some of those kind of agility around controls you're talking about, but of course you're making me think of you know, not only do we have the agent uh Gentec AI apocalypse coming, we have the you know, the quantum computing apocalypse on the crypto side of things, but we'll put a pin in that one for a minute and kind of come back. And I want to talk about some real world stories maybe that that will resonate with our audience uh and our listeners here around this, which is you know, and I think we were talking about some some anecdotes, including there was an HR story, right? So, you know, someone's using uh AI inside this. Uh refresh my memory on that one, man. What what was that?

SPEAKER_01 21:09

We had an instance at a customer where uh an employee was asking their co-pilot for a device about a situation with another employee, uh with one of their coworkers, and the co-pilot felt like it was appropriate to send an email to HR to uh help resolve that issue. Right. And so there's a lot of unintended consequences with AI, right, as we've seen with the hugging face incident. But like there's micro versions of that happening every single day at companies uh that like most people aren't paying attention to.

SPEAKER_02 21:43

Yeah, I I think it's interesting. It's not just like the headlines with the the hugging face and and other things like that. You know, we're definitely hearing kind of anecdotally and and confidentially people talking about red team scenarios, right? I point my frontier class agent, um, and in fact, I Um subject for another talk, but you know, I pointed some stuff at some open source when I had access to Fable, and it's amazing what this stuff can do. So let's say you have that, you point at a service, it finds a Salesforce plugin, it keeps on going. So, like in that scenario, kind of back to our whole responsibility thing, who's responsible? You know, what what happens in those scenarios?

SPEAKER_01 22:20

Yeah, the containment is a major issue right now. Like we have customers who are using um the frontier models to do penetration tests of internal systems, right? That like everybody has been subject to what's called patchmagedd recently, right? Which is like there's all these up, there's all these vulnerabilities that are being found by AI. And responsible businesses are saying, like, okay, how can we get in front of this? How can we point the frontier models that are lower-laying environments to find our vulnerabilities, fix them, patch them before somebody finds it from the outside? Um and they're very concerned about, hey, you know, how do we make sure that these agents don't start targeting internal production systems? How do we make sure that they don't um overwhelm the internal systems by in terms of traffic? How do we make sure they don't break out of this sandbox and start targeting our customers? Because, hey, you know, if if uh if we want to have a winning organization, maybe the only way to win is by taking down all of our customers' websites. So um that's uh that's a not ridiculous leap for an AI agent to make. Um and so what we're seeing companies start doing is uh attaching what we refer to as evaluator agents to the workflows of their pen testing agents, which is interesting.

SPEAKER_02 23:44

Interesting. Evaluators. Okay, yes. This this is a this is a term that seems like it's starting to get some currency, right?

SPEAKER_01 23:50

Yeah, which is like, hey, you know, if if I have an agent running a pen test that gets a specific internal service, I'm going to have an evaluator agent watching what it's doing. And I'm going to have it um alert my rec team leads if it starts going out of bounds. Right. And and so what we what we've seen is you know, the the pen testing agent is running a pen test. It's spun up 20 different subagents to create new vulnerability classes and things like that. Um and then there's a Teams channel that's happening in the background that's just keeping the red team humans up to date of like what's happening, how it's working. Oh, it started it made an API call to this production system. Let's stop it. You know, things like that.

SPEAKER_02 24:38

There you go. So making sure that there's still, you know, you heard these these different terms of human in the loop, human above the loop. I mean, I guess I I want to hear your thoughts on that, but before we get there, you know, it's like maybe hearkening back to your pen testing days, right? You know, it was always the case that you had to have a letter that showed your rules of engagement, especially if you're doing like a physical pen test and you're gonna be on site, you want something you can show to the IT operators in the data center or you know, the security guards. You know, it seems like some of these agents are just incented to win so much, you know, kind of seems like part of that is somehow we need to be able to put those guardrails on, and that evaluator agent is maybe a way to do that. Is that kind of one of the ways you're looking at that?

SPEAKER_01 25:25

Absolutely. And we think that you know companies will have multiple evaluator agents for every agentic session in their environment, right? So you may have an evaluator agent to make sure that you know your agents are only accessing the right type of data, or you know, that your finance team is staying in line with PCI compliance as they're using AI. Um, or you know, in this case, pen testing. So I think we're going to be using we're going to have multiple agentic eyes on the actions of most agents that are operating in our environment because we want to stay safe. And this is super important because you know an agent can operate a thousand times faster than a human, right? Like I'll give you an example. Like when we when we first implemented our internal Salesforce instance with uh agents, we noticed on the Salesforce MCP server that there was a uh a tool called run a SOQL query. And basically, like my agent can run any Salesforce query and and execute it on my behalf. Like for me to go into Salesforce and delete a thousand records, it would take my fingers a long time to do it. For my agent to do it, it can do it in a millisecond. Right. And so having these real real runtime controls with evaluators in place is going to be really critical because just because the velocity at which these things can move are something that we've never had before.

SPEAKER_02 27:03

Yeah, that that kind of shifts it to that the that one phrase I was saying, right? Human in the loop. Like for a long time we're saying, you know, I still want to have a human in the loop. With things moving that fast, you know, it's kind of like human in the loop versus human above the loop. I don't know. Anything you want to kind of comment on there, especially in these cases where you've got these swarms of agents, right?

SPEAKER_01 27:26

Yeah, I think you're gonna have like you know the loop, and then you're gonna have a set of agents above the loop, and then you're gonna have a human above that. Because the thing that we also want to avoid is overloading the human with decisions. And I think everybody went through this journey when they first started playing with Claud Code or something like that on their own computers. You know, it would ask you every prompt, like, can I have access to this? Can I have access to this? Can I have access to that? And then the first couple of hours, all of us read those messages and actually took it into account. Now cloud code shifts with auto mode by default because they realize that nobody actually reads any of that stuff anymore. So I think we have to be very judicious in when and how we engage humans, and it has to be relevant and timely. Otherwise, people just say yes.

SPEAKER_02 28:17

So that kind of shifts to this idea of um the best controls are invisible. You want the agents to actually help you do the work, right? You know, it's it's it's I don't know, maybe some of this is it felt like some of the agents we were supervising last year were kind of like you know high school interns, and some of them are now finally college interns or or early in their career. It's like, how do we kind of get that those controls in place so that if they actually start shifting away, you know, uh when they misbehave, what does that pattern look like? I mean, I guess maybe let's let's walk through one of those those patterns, which is you know, first of all, since you know hyper obviously has a deep identity background and and things like this, is the agent's um um identity tied to my identity? You know, what are the different kind of use cases around that one? And how do we put those right controls in place?

SPEAKER_01 29:12

Yeah, right now what most companies use agents, they're acting on behalf of humans, right? Or companies are taking workload agents and then assigning them to humans to be able to be used. But even in those instances, you know, when when if I have a workload agent that's in the cloud somewhere, a human is kicking off a workflow with that agent, which is like read my emails and summarize them. Like that's a really simple thing. But it's a human asking that AI to do that on their behalf. Um, so first thing is we want to assign the human with the right set of agents that builds the accountability in by default. And then two is you'll have a set of evaluator agents that make sure that those uh those AIs are staying on task, or at least keeping the company within compliance, right? Making sure that the agent isn't making some random API call or downloading some packages from places not supposed to, so on and so forth. Um, or ultimately what we use it for a lot is making sure that the agent is following our internal SDLC process. Um and then the the second thing you wanted to make sure that you do is be able to loop the human in only when appropriate. Um and that should be a non-deterministic system, as much as you can help it. And what we've seen is companies starting to layer their agentix security program in like an onion type of shape, which is agents have more autonomy for internal things rather than external things. So, like for us at Hyper, my agent can send emails to my coworkers, but it can't send emails to our customers. Right. So identifying those types of scopes, uh, which are very fine-grained, um, is super important. Uh, and I can share more about that too.

SPEAKER_02 31:04

Yeah, I'd love to love to hear more about that, but let's maybe pull in another kind of real example. Um, I think you know, one case you're talking about you know, a defense customer you have or someone who's working in the EU. And you know, this is really top of mind for a lot of folks with things like the Cyber Resiliency Act just going into effect earlier this month and more things around this, which is it's hard enough to make sure that you keep your data residency right with a group of humans. Kind of what does that look like? You know, in in that case where you know, hey, company with thousands of international employees, there's sensitive data that should not go to a US model if it's in the EU, and maybe likewise in in the US, they don't want it going to a Chinese-based model, you know. So what do they need? You know, what does this look like once you get it running?

SPEAKER_01 31:52

We have this, you know, government contractor, customer who operates in numerous countries and advises their government and and works with their uh intellectual property of those governments. And so it's very important for, let's say, the French government and the systems that are being used there to not make their way to the United States and vice versa. And so uh proper data classification is absolutely needed. But then how do you make sure that agents aren't sending this data across country lines, which is super critical? So for example, like a a French employee in France using an agent to view a French defense document, perfectly acceptable. But then that agent takes that data and shares with another agent that's running in the US to complete some task. How do you prevent that type of cross-pollination of data is super important. And so this is why it's very able very important to have those breadcrumbs that you can enforce in real time. Because if a French employee is accessing a French document, gives that document to their agent, and then the agent takes information for that document and sends it to an agent in the US, which it's not supposed to, that has that control has to be applied at runtime. So you really have to be able to chain the data custody across all the various tools that the AIs are using and how they're using them and enforce them in real time. And and the reason that the in-real time thing is so critical is because agents can do this really, really fast. Right? If for a human to like upload a document, download it from a system, and then try to move it into a different system, even if they have to like find a workaround for it, assuming that they're not like malicious in any way, it would still take them pain and effort, and they might not know how to do it at all. An agent can find those paths very easily.

SPEAKER_02 33:49

Yeah, it seems like our old school alert of something goes to the person themselves, say, hey, did you make a mistake around this, or to their manager or to the IT team, depending on kind of the company's policies? You know, those kind of detective controls that get into prevention are no longer sufficient, right? You really need to have some stuff that's going to operate and at least stop things, or you know, as best it can in real time.

SPEAKER_01 34:14

Yeah, look, an agent can go on Heroku, create an account, spin up a web app and an FTP server or something. Right. And then start uploading stuff there, and then telling another agent that's in the United States to download it from there. And unless you have like another AI looking at that, those actions, like a human would not know anything. Right. So uh it it's it's really fascinating and a little bit scary. And that's why I I'm just such a true believer and you can't have deterministic controls for these things, it just will never scale.

SPEAKER_02 34:48

Well, Boyan, I really appreciate you know you being very candid in this conversation. You know, candid CISO podcasts, generally we're having conversations with CISOs, and I know your CISO has lots of interesting stuff to say, but you know, I appreciate the depth you've been able to bring to this. So, and also, you know, not making it kind of be a hyper product pitch side, but I do want to ask some candid questions about this since you have been dealing with folks you know here and doing real stuff. You're not just kind of you know getting funding with your first uh you know uh slideshow and uh agentically designed uh you know uh minimal viable product. You know, you guys actually have some stuff out there. You know, a lot of vendors say we have design partners, we're doing stuff for their newest thing. Right now, you have some of the new stuff you're working on in production with two customers, and you're not announcing that yet. You know, our CISO listeners will appreciate this, both on the confidentiality. Obviously, we're not going to talk about those specific customers and the drive for real world cases. What kind of feedback are you getting from these early adopters?

SPEAKER_01 35:49

I mean, the the feedback is generally overwhelmingly positive, but I think from my CISO friends and and some of the peers I had when I was you know running a security team myself, we're in a we're in a phase of time right now where um collaboration is just more critical than ever.

unknown 36:09

Right.

SPEAKER_01 36:09

And we're in the like it's kind of like when cloud first started becoming a thing. Um, I felt that I felt like the security community really came together at that time, which is like, hey, my CEO is pushing me to get to the cloud so we can move faster. I'm tired of all this on-prem stuff. How can we what do we need to put in place to make it a success? And I've I'm getting that same um energy right now from the industry, but it's like a thousand times bigger. Um, and and and the pressure is is much, much more significant. So we've been getting amazing feedback, but I think um, and the reason that we haven't been super loud out in the community around like us having a product in the space and solving world hunger with it, whatever, um, is because like we're still learning ourselves too. Uh we don't claim to have all the answers by any means, but we know that we can solve the top you know eight to ten use cases extremely well. And if we just focus on those and and get many more customers solving those use cases with us, then I think we can expand from there.

SPEAKER_02 37:15

You know, a number of folks have kind of talked with me, and I'm sure you as well. I've heard a number of these conversations about it feels like a similar point in time for us when many organizations were moving first to the cloud. And you're this great transmission to kind of a plug for something we're doing together and you guys are helping to sponsor. So, you know, uh Jason Chan was one of the guys early on there at um Netflix and you know came up with this whole idea around kind of the idea of the paved role, paved roads, paved roads, not just kind of gates. It's not the department of no is you know, I IT for a long time was called the department of slow, I uh security the department of no. How do you turn that around and kind of you know uh um enable your business teams to get things done? So he came up with this whole idea about paved road, and that worked well because the platform team made that safe path the easy path. And you know, identity made it pavable, right? So who paves that road for agents? And can it still be optional when the driver is a machine, right? Because we used to say on the paved road, hey, if you really want to go through the bush and do your own thing, you can do that. It's gonna take you longer, it's gonna cost more. You know, in an agentic world, you know, are do we have to stay on that paved road, or you know, what does that look like in your opinion?

SPEAKER_01 38:34

I think we have to get really good at paving roads extremely quickly because uh the the speed at which an agent can work, you know. Previously, if you needed to spin up a cloud system, so on and so forth, like you needed technical experts to do it. And it took time, took months, weeks of planning sometimes. Like now with AI, a non-technical person can spit up, spin up a web service and start doing stuff. Right. And so that paved road has to, uh, you know, I hate I hate to use the word like shift left, but like we have to provide the capabilities to the non-technical people out of the box to be able to do this. And so I think one of the best things we can do as a security community is go internally within our organizations and show them how this stuff can be done quickly and safely and lead from the front. Um, I think if we can show our marketing teams, our finance team, our sales teams, how they can leverage agents very quickly and very reliably using company-approved tools, I think we'll be in a much better spot a year from now.

SPEAKER_02 39:42

No, I think I think you're you're spot on there. And and it's reminding me of a of a real world scenario from for me in the past, like in this one case where we had uh uh the startup CEO and and founder said, we're gonna go down this path. And I said, that's great, because I'm tired of trying to tell, especially some of the more technical folks, no, you can't use open claw with a totally you know open harness and get all these things done. Here's where we have we have some minimal controls here, there's gonna be more. And that way you could have that conversation with the finance department, and they're saying, Is it really okay for me to point this at some spreadsheets? And I was like, Yeah, within this environment. Now it's the same kind of thing, it's a trust but verify scenario. If you see something, say something. So, you know, uh accounting lead, if something seems like it's wrong, please do talk with us and we'll kind of see what's going on there. But yeah, this whole idea of we have to do this machining in such a way that you know everyone can use it, not just our our techie folks, which was kind of what happened in that first you know, uh set of things where you're building it for just the engineering department or the product department, other things like that. So, well, I will put in that plug and say, hey, that's the question that Candid CISO Live is going to answer on October 8th at Stanford this year. So this podcast will be out if you're hearing this. Um, we'll have a registration uh link in inside the the show notes uh for this. And Boyan should be there as well. I think he's got a conference up in San Francisco, so we'll get him down the peninsula to Palo Alto, and you know, we can continue the candid conversation, not just with Jason, on kind of his thoughts about what the paved road looks like in this new era of agentic everything, but you know, some real world cases where you know Boyan's been working with folks on this. So I'll say uh join us there. And before we sign off, I'm gonna give Boyan uh you the kind of the last word here if you'd like. Anything you'd like to make sure our listeners are aware of, kind of from the learnings you've had of this kind of whole journey from identity for humans to agents and and the mix of both.

SPEAKER_01 41:46

Yeah, I think um don't try to reinvent the wheel, is what I would say. Um, and and really use the tools that are that you might already have in-house to implement some of these baseline controls and really just think about the accountability model and and how you can maintain a similar accountability model to what you have today with agents and the least, the least resistant path to get there. So, yeah, I look forward to being there on October 8th at Stanford. Um, those of you who want to learn more about hyper, check it out on hyper.com slash candid CISO. We'll have some material up there. Yeah, I look forward to seeing you guys in a couple weeks.

SPEAKER_02 42:23

Awesome. Man, I feel like we could keep going for hours here, but I know you're a busy CEO and uh you've got your next thing coming up. So I'm just gonna say to all our listeners, thank you for showing up for uh our inaugural session of Candid CISO Season 2. Look for great conversations, both our in person one on October 8th, and more candid conversations coming up in the season,