エンタープライズリスクに向けたAI対応コントロールの構築 ― Hyperproof創業者 Craig Unger 氏との対話

Welcome, everyone, to the Emerge AI in Business podcast. Today's guest is Craig Unger, CEO and founder at Hyperproof. Hyperproof provides a platform for managing risk, compliance, and security work across large organizations. In this episode, Craig explains how AI is adding new layers of risk inside the enterprise, from machine identities that now far outnumber human users to unapproved AI tools and vendors building AI into products companies already rely on. He also discusses how AI can help risk and compliance teams scan for regulatory change, test complete sets of evidence rather than samples, and support dollar-based risk modeling while human experts continue to define the controls. Today's episode is sponsored by Hyperproof. In this episode, we cover how AI is changing enterprise risk. To go deeper on this topic and learn how leading organizations approach AI investment more like a venture portfolio

and why interdisciplinary collaboration is critical to defining the right data for AI success, download our free PDF report beginning with AI at emerge.com slash AIK1. That's E-M-E-R-J dot com slash AIK1 to download your copy. Now the conversation with Craig. So, Craig, welcome back to the studio. Thanks, Yolanda. Great to be here. Last time you were here, we talked about how fast rules are moving. And then today I want us to look at what's changing inside the company. AI tools are showing up everywhere. Some of them are approved, unfortunately. Some of them are not. Vendors are adding AI to the products that we are already using.

And some of those tools aren't even just making suggestions anymore. They're taking action, which is scary and exciting at the same time. So how has AI changed the GRC landscape? What are the crucial new security threats, risks, and behavior that our GRC leaders are concerned with? Yeah, Yolani, great question. Before I get into the GRC pieces of it, I will just talk for a second about the fact that the AI itself, in terms of a vector for security infiltration and breach, I don't want to understate that. I mean, that's actually kind of an enormous concern for the more security-oriented vendors. I'm not going to touch on it a ton today because we're going to look at kind of the governance, the compliance, and the risk pieces of it. But that's obviously already out there, and it's keeping all of my kind of compatriots who work on the security products very, very busy. But in terms of the products that we work on, I guess, you know, a lot of our workflows have to do with how you control, how you create policy, how you enforce that policy inside your organization, right? And I think one thing that's been going on for a long time that has actually been accelerated by the AI is this concept of identity proliferation.

I mean, the identity and access management area has been a major growing area in the software space for a while. But when you think about it, you know, there's been kind of a, let's say, individual or human being present form of identity where you log in, you're challenged, you're there. And then there's also scenarios where you're not present, where you're actually presenting some type of API key or a token or something else in order to get access to a system. And where some listeners might think, well, that's actually the problem in and of itself. That's actually not really the problem. that it's very, very important to have those kind of modalities of access because that's the way that automation gets done, right? That's the way that you're able to write some code, and the code goes and gets something done for you when you're not there. So we're not looking to shut that down in any way. But the issue is that some of the methods for being able to get access to systems when a human isn't present are the exact same modalities that an AI uses to get access to those systems, right? So you think of, like, again, tokens, API keys, and those things, And there was an interesting study that was done recently by Palo Alto Networks, a 2026 identity security landscape study that even surprised me.

And I'm in the space. It showed that there's 109 machine identities per human identity in an average enterprise, 109 machine identities. And, of course, you can imagine that a lot of the software to control who accesses what system are very much built around the idea that that human being is on the other end of it. It's a real human identity. They're logging in. They're doing those things. And when you have these kind of long-running or long-lived keys and tokens that are just out there for a longer-range type of access, it becomes very, very difficult to understand how to govern that because when those things are created, they're not really created along with a very firm purpose for why they're created. So some of them can get used for all kinds of different reasons, some of which may be what you intended. When you allow that, tokens get created and some others not. So I think back to that kind of the first problem is just this idea that, you know, the systems that we're all trying to secure and assure are compliant now are getting hit by 100 or more different identities from one individual without a real system to be able to manage that.

So that's the first thing that I would hit. Building on that, right, not separate, but building on that is this concept of shadow AI. And shadow AI is the use of the AI tools and models inside an organization in order to give the users some type of, like, visibility to certain assets that you may not have thought that you were getting them access to. That's the shadow part. You didn't actually intend for that to happen. There's a couple different ways to think about how to improve that or where you might be able to stop that. One would be on the data store side, so where the data is actually getting pulled from. or the database or whatever the live system is. And then the second one would be where the request is really being made. And the problem is that, like, neither of those are really that clear how to, you know, lock down. Because on the data store, it's not really always that clear to whatever access management system, as I was saying before, whether or not this particular transaction is supposed to be allowed.

It's just not that easy. You can't just go and say, well, any kind of AI, you know, type of transaction that wasn't contemplated before, the token can no longer be used for it You just can do that So the data store side is difficult On the request side it difficult for a number of reasons The AI and prompting is relatively new for all of us But also, you just kind of never know what people are going to prompt and how an AI is going to respond to and how an AI is going to take that apart into different transactions. And also, even if you could kind of look at a single transaction or a single prompt that's going on when a user types it, that one may be fine. And it may be the fact that the user is doing a composition of five or six different requests that you kind of hadn't contemplated, that the user may be able to do all those things at once and get access to a certain system or a certain kind of set of data that way. So, you know, I just think that it's kind of a different thing. There's a deeper set of actions the AI can take, compose it in different ways. So, you know, I guess if it were easy, we wouldn't be talking about it.

It would be a solved problem. But these are the challenges that are being layered on top of the systems that people have already had to secure. What is your thoughts about vendors adding AI to the products that you're already using? Like, how do we know what they're doing with our data? Well, that's actually a really good question. I mean, there's not also a single answer to that. There's really a series of different things that you have to watch out for. First and foremost, you want to be able to look at what's going on in the terms of use or the licensing agreement that you're signing with these organizations. All that stuff needs to be, you know, kind of laid out in terms of how data is being used. I'd say even especially if you're in Europe, because there's even more kind of a higher bar to what has to be described in terms of how data is used and the ability to be able to kind of claw back that data, the right to be deleted, all of those different things. But beyond that, there's also other sources of what the evidence is in terms of how the vendor is actually using your data. And a big one is what we do every day, which is what type of compliance frameworks and what type of governance are they putting into their systems? And then there's a whole kind of different set of modalities for how to check that.

Sometimes it just starts with asking them during the sales cycle, like, hey, what is it that you exactly do? And how can you document it for me? Other times it could go to the next level of saying, hey, I want to see what you do and your policies, either in some form of a trust center that I can go look at, or I want access to your compliance documentation. It's quite common for a vendor to, like, share their SOC reports or whatever, their ISO reports, so that you can take a look and say, well, it met their criteria, but does it meet our criteria? And then, of course, on top of that, you can layer special contractual obligations, which exist in different frameworks. Maybe it's kind of a data processing agreement, as folks might be familiar with in the GDPR, or, you know, whatever it is, the different frameworks have some different capabilities for that. And it can even go to actually automated scanning, right? there's systems out there that will scan to see how certain data is dealt with, be able to try to ascertain what the policies are that are actually being implemented by the vendors. So that whole area, which broadly falls under third-party risk, we at Hyperproof have a product for that. I didn't even mention the questionnaire capability.

I don't mention it because it's in our product, but it's kind of the way that the industry has kind of come from, the place that the industry has come from, but there's some of the also more interesting places that the industry is going toward. The crazy thing about technology is that we just dive into all the risks that comes with it. But then on the other side, that same technology that creates the risk also helps us to manage it. I keep thinking about the board, and after our last conversation where we established that, the board does not want red, yellow, or green charts anymore. They want to know exactly what a risk means for the business and whether it's under control, and that's where the focus is. So where are the places that AI-enabled workflows improve risk visibility, reduce compliance friction, or even strengthen control assurance? Yeah, and the timing of that question is really interesting, given that we just talked about how we're all nervous about what these vendors are doing with their AI. But, of course, we want to be able to leverage our AI to make sure that things are more safe and more secure, right? And that's kind of the way it works out there. But I've been saying for the last while, you know, with AI, it's so powerful in so many dimensions.

But for me, the really standout is the ability to take large amounts of data and be able to really kind of analyze that. I feel like today's AI is a lot better at doing that than it is at actually making decisions. You know, it kind of like takes the data, analyzes it, and then surfaces up some possible decisions that you can take depending upon which inputs or assumptions are most important to you. And I think that those challenges of being able to get access and process large amounts of data actually are challenges that have plagued the compliance and risk area. And that's why the AI tooling that we can add to our applications is so helpful. So I'll give some examples. So one example would be horizon scanning, what's called horizon scanning. And that's the ability to kind of look across the web and know what's changed in terms of regulations. You know, some country, you know, let's say has now like published some new privacy policy that, you know, you have to abide by. Like, how would you find that? It's like a needle in a haystack. And so, you know, in the old days where there had been almost like subscriptions to systems, maybe it'd be like Nexus or, you know, or Thomson Reuters or whomever.

And those still have their place for sure. The AI can actually make a lot of that stuff much more just in time to know, hey, there's actually change that I actually have to respond to today. So that whole thing on horizon scanning and getting access to all of the kind of unstructured data, meaning the policies and regulations that are out there and how they've changed, that's one modality that I think is really easy for us to apply AI to and something that's a real obvious benefit. Another one that I would talk about is evidence collection, right? The heart of the compliance workflow has been collecting evidence in order to assure that your systems are working the way you want. Oftentimes that's who has access to what system, what type of data is getting written to those systems, you know, those kinds of things. Obviously, the AI has tremendous ability to be able to access that data and be able to analyze it and give us the kind of data that we want in a much more quick and more up-to-date way. And I think that's an important part of the AI. Typically, the evidence is tested after that.

And so there's a lot of different forms of testing. I mean, one interesting part of it is that the audit industry has grown up knowing that the data is difficult to get in kind of its entirety. And so there's this whole methodology of sampling the data to make sure. If you sample enough data points, you'll know whether the organization is kind of doing what it says it's doing. but increasingly with the AI, you don't really need to sample the data. You could actually test the whole population. So that kind of one interesting change in terms of how testing is approached given the existence of AI But another one would also be the type of testing A lot of the testing that gone on historically has been the testing that very easy for machines to do. So that's comparing a value to another value, checking to make sure an individual who shouldn't have access didn't have access, all of those kinds of things which tend to be kind of true or false type tests that are able to be expressed with logical predicates. You know, if this is greater than that, and this, blah, blah, blah, blah.

But things that have been harder have been how do you look at, like, a long document or set of documents and see how is the information linked? How do people navigate from one place to another? How does the context of the information in this section of the document relate to a policy that may be somewhere else in that document or in another document entirely? So what I would think about there is the ability to test what has been typically talked about as non-structured data. You may think it was a PDF or WordDoc or whatever, an HTML file itself, you know, the ability to test your webpages. All of those things have gotten a lot more approachable in terms of being able to not only have the AI do the testing, but also have conversational AI be able to engage you so that you can describe the test without having to do anything that looks like writing code. Right? So you can describe those tests in English. Right? And so I think those are some of the keys where the AI is really playing well today. I think in the future, one thing I would call out is we're really interested in this idea of how to leverage the AI for risk prediction.

So when you have access to that really large amount of data and you can look at it a lot of different ways, there's a lot of interesting ways you can model the risk and give the user the ability to kind of run those simulations that tell us where the risk is in their org. And that's where the AI really kind of starts to ride alongside you and really helps you reduce the risk in your organization. So if we're thinking about AI for risk prediction and getting us to reduce that risk, do we also see the AI getting involved in setting the controls for the risk? You know, I think the controls themselves, they can be suggested by data. They can even do that today. But I think at the end of the day, that's an area where the body of data you're looking at in terms of the controls themselves is small enough that I think you're going to want to have human beings who tend to be like security experts and those things come up with those protocols. Maybe an interesting way to kind of think about it would be let's take a totally different area like medicine, right? You could have the AI diagnose or even do a procedure, but when it comes to actually defining what really the problems are and the parameters are and how that really would play into how long you might live or what your quality of life might be, those are areas where you probably want the experts to be.

And then, like, the more kind of repeatable parts of how do I find it and what do I do about it, that's probably where you put the AI is what I would think is going to happen. So you're always going to have your security experts and, you know, especially when you think that you have a whole set of human beings around the planet that are just kind of gulling themselves with outsmarting the AI, right? And so those are the same people that you want in some kind of white hat type of way to be able to help you define what to look for, right? You don't want to leave that to the AI itself. It might be one level too self-referential. Yeah, I like that. I like the picture that you painted there. I think that is going to definitely line with the audience. They'll see it. So the AI is extremely smart and the expert in some areas, but for certain decisions, and I guess where dollars are on the line and lives, obviously, that's where we want humans to be involved. And that's where we want to also, you know, see the risk in dollars. And on that topic, this is also something that we touched in in the last episode, you know, risk in dollars, imagining that. How is that dollar amount actually calculated? Do we use AI to calculate that while it's calculating the risk?

Is it part of that workflow? You know, it's interesting. The current systems that are most used came out so far before AI that they didn't use AI as it is thought of today. They used more of like a simulation modeling, like a Monte Carlo type simulation, which certainly is not really an AI in and of itself. However, what we found is that when you combine an AI up front to be able to make the connections between the data elements and the controls of policies, perhaps suggest some different attack vectors and some other scenarios that represent risk, and then combine it with the tried and true ways of doing this kind of dollars and cents methodology on risk, That's very, very powerful because what it does is it gives you a way to get coverage of your risk universe that was just like too manual to have to do before. That's where we're seeing a lot of it. In fact, what we're looking at is like how do we really kind of build that in to our products in a way where the users have to have so much sophistication in what the risks could be because those are pretty well understood out there.

And really, the AI can help them take the important information about their organization, their industry, where they are in their maturity journey, and all those things, and then help them make those connections and then smoothly lead them to the risk modeling. That's really where we're applying AI in that scenario right now. Well, that's interesting, and it makes a lot of sense. So, Craig, let's bring this home. I have a picture of a risk leader listening to this on their commute to work this morning. They agree that AI is changing their risk. Kerry, there is a lot of work to do here. So what steps can enterprise leaders take right now to apply lessons and begin modernizing their own GRC workflows in what would be a measurable and operationally meaningful way? Okay, so I'll hit upon a couple of things. Some touch on where the AI is the risk and some touch on where the AI is actually the benefit in actually applying it. In the former case, I think making sure that you have a set of policies and some type of inventory for where your AIs are running and what they're doing is very, very important, right?

There's AIs you know in terms of people getting access through, you know, the browser and UI. And then, of course, there's all the programmatic access that happens to the different AI systems. And being able to kind of lay out, here's what you want to use these AI systems for. Of course, you can kind of carry it further than that, which is what kind of risk does it present and some of those other things, and getting the kind of right governance model around it. So I'll call that governance and identification of it. The identification of it is it's an area that's moving fast. I mean, you can start out by trying to lay out in type of list format and have some type of more manual process for how people register where they're using it. But pretty quickly you going to want to have an automated system that goes and uncovers right what type of AI is being used and what type of risk factors that that kind of throwing out there right So and that even before you have anything automated that can do anything about it It just more to understand what there So I think that like probably the key of how to really at least get started on it now The things I would recommend going beyond that are going to radically change in the next six months and nine months.

It's kind of more tooling comes on board to actually take action against it, but at least start inventorying it, understanding it, understanding how you're going to operationalize it, and putting a model around it. Now, in terms of where the AI is also helping, I want to go to a second piece, which is, you know, automating this kind of continuous control monitoring that we've talked about before, we did it here as well, is like a crucial thing. Like you want to make sure that you can actually use the AI to grab all the data that you need and continuously look at the evidence as we've discussed it and run the tests that you need. We are so far beyond doing that once a quarter or in preparation for an audit. It has to happen consistently. It has to be running consistently. And then when new systems come on board in your company, they've got to be added along some kind of SLA. That has to happen. That's been, I would say, a trend the last four or five or six years. But now with, like, just this kind of flowering of the kind of AI tooling going on, you really can't afford to have it be manual anymore.

It has to be automatic, and it has to be kind of continuously monitored. Relative to what you said, Yolandi, coming back around, you want to marry some of the first two things that we talked about by looking at your controls again and making sure that they are now AI-aware controls, right? So a good example would be in a company that may be doing SOC to, let's say, I'm just going to pick a random framework. They might have a couple hundred controls, just to give you a sense of scale, for what it might take to become SOC compliant. Not every one of those is going to be affected by the fact that you've got AI. I don't even know how many. Maybe it's 100 of them. Who knows? But some set of them are going to be affected. And you have to widen the purview that you're looking at each control through now that you have this AI tooling. And then you have to go through and reconnect it in a continuous way to make sure it's monitored. So now you're actually bringing it all together in terms of where is the AI, what controls of mine does it affect, and how do I automate that? I think those three steps are a pretty good set of steps. Identifying where your AI is, doing a continuous monitoring system,

and then connecting the first two things by widening the purview of your controls into where your AIs are running. Those are some good steps. The last thing I would add is kind of more of an external perspective, which is what are your vendors doing? Because more than half the breaches that happen to organizations, they happen through their vendors. We probably talked about that before. So there's a lot of modalities for how to go to your vendors and figure out what it is that they're doing. And we've actually interestingly hit upon some of them here in response to another question you asked. You can actually look at their compliance frameworks. You can actually go and examine your contract. You can look at their terms of service and the licensing terms. Of course, you can go to the questionnaire-based approach. What are you doing, right? And all those things. There's automated tools. So these are things we've already talked about in this conversation. But bring that into the fold and make sure that you're taking those approaches to understanding what your vendors are doing and layering what it is that they're doing in AI so that you understand how they're tackling that new part of the problem externally the same way you're trying to understand how you're tackling it internally.

Absolutely. This has been great. I think my message for the audience today is definitely that AI brings new risk into the company through things like tools that nobody approved or that some of the people aren't even aware that they're using, sometimes through your vendors like we just mentioned, And sometimes because the AI acts now and does not just make suggestions anymore, but that same AI idea can also reduce your risk. It can also be tied to controls that help you to bring down the dollar value of that risk. That is the ultimate message of this conversation. Craig, any last thing that you want to leave our audience with today? I mean, just a little bit of, like, motivational words and the fact that, like, this is the biggest change, right, that I've seen through my career. It's no surprise if you feel that you're a little bit caught off guard with this. And what I found in most things in my life, and this is no different, is that, you know, you finish faster if you start earlier, right? It's nothing more than that. You don't have to wait to have the whole solution to get started. Take your baby steps. You know, it's not crazy to just get a set of people in a room to actually discuss this and come up with how you want to tackle this,

even if that seems like a very kind of low-tech solution to start getting started. once you have that kind of consensus of what you want to do. The technical solutions and automation opportunities and all those things will start to present themselves. But none of this stuff is just supposed to pop out. It's all supposed to come from the results of a very reasoned set of individuals having conversations about how they want their organization to run. And so don't hesitate to have those conversations knowing that it's the first step of a journey and you don't know exactly where that journey is going to take. Great message to leave our audience members up. Craig, it's been lovely having you on the show again. Thank you so much for coming back. Yolani, thanks very much. It was really fun to be able to talk about this. Wrapping up today's episode, I think there are three key takeaways from our conversation with Craig. First, organizations should start by inventorying

where AI is running and what it is doing. Then put a governance model around it, moving from a manual list to automated discovery as tools mature. Second, AI makes it possible to monitor controls continuously and test entire steps of evidence rather than samples, so control checks can no longer wait for the quarterly review or audit preparation. And finally, existing controls need to be made AI aware, and the same scrutiny must extend to vendors that are building AI into the product the organization depends on. In this episode, we covered how AI is changing enterprise risk. To go deeper on this topic and learn how leading organizations approach AI investment more like a venture portfolio and why interdisciplinary collaboration is critical to defining the right data for AI success, download our free PDF report, beginningwithai at emerge.com-aik1. That's emerj.com-aik1 to download your copy. On behalf of the team at Emerge, We'll see you on the next episode.