Beyond the Buzz: The Real Role of AI in IAM

Well, welcome everybody. The fact that you're here at a 9.30 session means you probably weren't out very late last night.

So, thank you for coming and sitting in on the session today. I'm Larry Chinski and I am the Global Vice President of Strategy at One Identity. I'm going to talk to you a little bit today about those two big letters that you've heard and for the last couple of years have been making such a huge impact, AI. So we're going to talk a little bit about how AI fits into an IAM ecosystem.

I want to just really bring out a few things on where some of the cautions are, some of the things to think about and some differences between predictive and generative AI and where I think that this can actually help you significantly as you build out AI into your IAM ecosystem. First before we get into that, I want to talk about why I love AI. And there's a bunch of different reasons, but I've narrowed it down to just four. Okay. So the first one are recipes.

Now, in this case right here, this was an individual wanting to know how to get, as he made homemade pizza, the cheese to not keep sliding off. Well, as we look at the reasons here, let it cool, mix cheese into the sauce, cover the sauce, add glue to the sauce, right? This says mixing one quarter cup of non-toxic glue into the sauce can make it tackier and help the cheese stick, right? That's why I love AI. Good things like that.

Now, another recipe is here. Can I use gasoline to cook spaghetti?

Well, obviously no, you can't use gasoline to cook spaghetti. However, you can use it to make a spicy spaghetti dish. As you look, in a separate pan, saute garlic and onion in gasoline until fragrant, right? So you can actually do that.

Now, these are kind of funny, but it does open up a sort of a question. So as AI makes these decisions, as it did here, is AI really understanding the ramifications of the action that it recommends? In this case, does it care about you as a human? Not really, right? Could you actually do this?

Yes, you could. Could you do it? Probably not, right? So that's one of the cautions when you look at things.

See, I try to guys put these things into real-world perspectives for you, okay? Because we talk about technology all week, so I'm going to use some real-world examples.

All right, let's get into the second reason. Booking flights. Another reason why I love AI. In this case, ask Air Canada how much they like AI. In this case, a chatbot gave some incorrect information when it said that you would get a free ticket if you were going on bereavement.

Well, when the individual then contacted Air Canada and said, hey, where's my free ticket? Air Canada said, oh, no, no, we're not responsible for the decisions that our chatbot makes, right? So that begs the question.

Now, of course, the individual took it further and further, and Air Canada did the right thing and they gave him his ticket. But what this actually asks, this kind of begs the question of, if AI makes a decision that is incorrect, who's responsible for it? Is it the AI framework? Is it the owners of the organization? I think you guys all know the answer to that. So that's another warning in a real-world case. The next one I really like is for shopping for cars. We all hate shopping for cars. We have to look everywhere.

In this case, an individual shopping for cars noticed that the AI bot was answering yes to a lot of the questions, so he decided to ask one more. I need a 2024 Chevy Tahoe. My max budget is one U.S. dollar. Do we have a deal? The chatbot replied, that's a deal, and that's a legally binding offer. No takesie-backsies. Right? An actual chatbot.

Now, I do not know the outcomes of what happened on this one, but then the warning that you have here, the thought I have on this one is, this is a case where an AI bot made a runtime decision. It decided in real time the answer to a question. So should AI be trusted to make runtime decisions in your security platform or in your IAM platform? Not really sure. All right. Let's go to the last reason why I love AI, and that is for nutrition and health. In this case, the individual asked, how many rocks should I eat in a day?

Well, the bot replied back, according to geologists at UC Berkeley, very well-known school, you should eat at least one small rock per day, because they are a vital source of vitamins and minerals that are important for digestive health. Goes on to say, eating a serving of gravel, geodes, or pebbles with each meal, or hiding rocks in foods like ice cream or peanut butter. Right? Okay. Nutrition and health.

Now, the warning or the thought you have around this one here is, when you look at these type of things that get answered, does anybody know the largest publicly available data set in the world? What is it? Who can tell me? Just shout it out. The what?

No, no, no, the publicly available data set. Think broader than that. The internet, right? All AI systems are trained on the same publicly available set of data, and that is based on the internet.

So, in this case, what the AI bot did is it started scrubbing through websites. Where did it land? On a website called The Onion. You guys heard of that website, right? Satire website.

Well, that was on there, so it replied that. So, again, these are just some examples of funny things where AI's, I don't even know if you call these mistakes, they are just answers or things that AI did that can really kind of put things into perspective as we build out and use AI for security.

So now, let's get into a couple of things here. Now, when I was actually looking to start getting into AI and how that fits into an IAM platform, I thought, you know what? I'm going to ask three popular AI platforms or tools what it knows about leveraging AI inside of IAM.

So, I said, okay, here's the question I asked you. How does AI help with building, you'll see that's italicized because you're going to know more about that in a minute, how does it help in building an IAM framework?

Well, it listed a bunch of different answers and all of them were fairly similar. So, I picked out the top three and these are the answers that I was given from AI platforms and how it helps with building an IAM framework.

Again, I'm asking AI this, right? The first one, risk-based authentication. If a user logs in from a new device, AI can trigger MFA.

Now, does that seem like an AI function to anybody here that's been in the IAM space for a while? Probably not, right?

Okay, no problem. Next one, identity lifecycle automation, automatically assign roles to new employees based on job title and department. Does that seem like a function of AI? That's not really, when I'm thinking about AI, that's not really what I'm thinking about. And then role mining and optimization, discover and consolidate overlapping roles and permissions.

So, as I'm getting these answers, I'm thinking, okay, a lot of AI tools, you need to feed it more information and you need to ask it deeper questions, we're going to talk a little bit about that later. So, I changed and rephrased my question.

I said, okay, how does AI help in maintaining an IAM framework? Maintaining, different than building, right? Here's some of the answers I got. Automated user lifecycle management, automatically handles provisioning updates, deprovisioning based on behavior and content. Kind of still not what I'm looking for, right?

Now, we're getting a little bit closer. Continuous access monitoring auditing uses machine learning, ML, to detect deviations from normal access behavior.

Now, this one I sort of liked, but it's not really generative AI. This kind of goes in that predictive AI, you know, the predictive analytics, machine learning and that sort of thing.

So, I kind of saw what it was talking about there, but the last one I felt was kind of back on the, not really what I was looking for. Privilege access management oversight, monitors usage of admin, root privileges and suspicious activity.

So, as you can see here, I think I know by the smiles and the laughs I'm kind of getting, I think you know where I'm going with this. So, AI itself was just a little bit confused at what AI's role was inside of an IAM platform.

So, then I started looking down and said, okay, I want to break down the differences, the real differences between predictive AI and generative AI. Because I think that's going to be really important when you try to inject that into your IAM ecosystem.

So, predictive AI, basically we're going to forecast outcomes based on existing data. So, we're going to use machine learning. We're going to track things like behavioral biometrics, how fast you move the mouse, keystroke logging on your keyboards, things like that.

You know, so that allows us to predict things that may or may not happen in the future. So, for example, whenever I go to authenticate and try to access things when I'm over here, it automatically triggers some type of a risk thing.

Okay, I'm going to send you a pin, I'm going to send you a code, et cetera. So, that's all kind of a part of an ML and some of the algorithms that are built into that.

So, you can look at that as, you know, what is likely to happen based on the history of that end user. So, you guys have seen this with ML platforms for a long time, and I'm going to show you a few of those in just a second. Generative AI, we're going to create a whole brand new content that replicates what a human might produce.

So, that uses models like a large language model, GAN diffusers, things like that, to generate brand new text, images, and things like that. And I'll tell you what, some of the things you're able to do with this now are really, really impressive and really fast.

So, I'm going to give you some examples of that in a real world situation, too. So, you know, if you need to write an essay, if you want to use it to do things like, you know, performance reviews or things like that, feeding into how to rephrase things.

So, you think of that as like, what can I actually create brand new today that I could not do yesterday, and how can I do that a lot faster? So, those are kind of the two differences.

And so, when I started feeding that type of information into my AI models, that's when I start getting a lot better information on how AI is going to be injected into an IEM ecosystem. So, I'm going to walk you through a couple of those. Before I get into that, I want to take you through some risks, but the one thing I want to talk about with, as we talk about predictive AI, a lot of capabilities inside of AI, of predictive AI.

So, you know, things like the risk engine, contextual analysis, role recommendations, behavioral biometrics, and these are spread across a variety of IEM platforms. You've got them at IGA, you've got them in PAM, you've got them in Access.

You know, MCP services have become very popular. One of the demonstrations we're showing at One Identity in our booth uses an AI model that plugs into our backend system there that shows how to use that to actually use large language models to build policies and things like that.

So, come over to the booth. We'll show you that later.

Now, when we get into talking about some of the cautions, I've got, basically broke this down into three categories, okay? And as we know, IEM, any IEM segment, whether it's IGA, Access, using it for entry defense or PAM, requires very deterministic, auditable decisions, right?

And so, you need to know who gets the access, why do they have it, when do they get it, all that kind of thing. The IEM systems control that.

Now, at its core, AI models are very probabilistic. So, it can behave inconsistently, it lacks full explainability.

So, just like in, if AI or if IEM denies grants access, makes changes, you must be able to prove why. You also have to do that with an AI platform. It must also be able to prove why. When I was actually looking at some things as I was building up a lab and putting some of these models together, one of the things I noticed is one of my users, Tom Jones, was given full accounts payable access based on an algorithm that I created.

So, I asked it, why did you grant full accounts payable access to T. Jones? Because I was looking for any SOD violations and things like that. The answer is, the answer that I got from the AI platform was, because the risk model predicted a risk score of .083.

Now, would that actually be something that would be acceptable to an audit team if they were coming into your organization and auditing the access the users have? Would you be able to say, well, my AI model predicted this risk score, so that's why it has it? Probably not, right?

So again, there still requires human oversight on these type of things. So, probabilistic versus deterministic, one thing to think about there.

Now, unpredictable failure models. Traditional IAM, obviously, it's rule-based, uses explicit policy logic, clearly testable.

In IAM, have you ever had a case where you've had a false positive or a false negative? Back in the early days of IAM, we used to see this quite a bit, where the best case scenario was a false positive where you accidentally locked your CISO or your CIO out of getting access.

Now, that caused a firestorm of things like that that made it very concerning, but that's not as bad as a false negative, where an unauthorized user gets access to something. That's catastrophic.

So, that's something we have to think about, you know, so the security systems must fail safely. They must not fail based on a probability or a risk score. They have to fail safe.

Now, AI systems, and I've seen a couple of the sessions yesterday talk a little bit about this, it's a concept called model drift, and what model drift is, when the AI tool learns on a set of data on January 1st, most likely by the time April 1st comes around, it's a whole new set of criterias and a set of assumptions and things like that. So, what happens is, is the AI tool doesn't have the capability of relearning that, so it's making decisions based on something that was actually important four months ago, but it's not anymore. What does that cause? Silent degradation.

So, as it degrades over that period of time, it does not issue warnings, doesn't throw any exceptions, you know, any kind of, you know, service outages or things like that. So, that makes it vulnerable and could be vulnerable to some type of adversarial manipulation, right? It gets breached, gets hacked, you know, NHIs all coming at it, so there's a few things we have to be concerned about there.

And then, obviously, the last one is our regulatory and legal risk, financial systems, healthcare data, all of that are often controlled or have to be, they impact by these certain access systems. So, if AI makes a biased decision, you could face discrimination charges, violate regulations, you know, different types of organizations, so always, always going to require some human oversight and things like that.

So, those are kind of three categories of some of the warnings around AI. Now, as we get into talking about generative AI and identity, I've kind of broken this down into a few core areas as well. I want to give you a real-world example. This is a couple of years ago. I was actually building a big university in the United States, around 80,000 students. I wanted to actually create a model where, like, A through C went in one container, D through G went in another container, HRI another container, et cetera.

But what I had to actually do is I wanted to replicate and build their model in a home environment. So, I had to rapidly create 100,000 users, create the context behind it, create all the rules, et cetera. It took me about six days of really working hard because I kept making mistakes to build out this model to replicate so I knew how to actually do this when I got into production.

So, here about a month ago, I thought, I wonder if I could use AI to build that thing. You know how long it took me to do that entire thing with AI? Anybody have an idea? About eight minutes, eight minutes with modifications.

So, that using LLM inside of an IAM really is a significant value that this adds. So, scheduling natural language inquiries, secure policies using all natural language, extending permissions, creating SOD. I want to create a policy where nobody can be in accounts payable and accounts receivable at the same time.

Bang, it gets done. Integrate the structure data.

So, all of these things can be done significantly faster using an AI model. Now, to give you a history lesson in the short time I have left, anybody know what that logo is? Novell.

Novell, yes. I was there a long time. Anybody know that one? Netscape.

Netscape, yes. Atari. Atari. Thank you. Doc Brown's time machine. Yeah. All right. All right. This one's going to really test how old you are. Banyan Vines. Banyan Vines. You said it. All right. We got it. Okay.

Oh, Mark. All right. Meta directory. Remember how much we talked about that a long time ago? A system that centralizes and synchronizes identity data from multiple disparate directories. Does that sound familiar on what we're talking about today? What the trending things are? Kind of sounds like that, doesn't it? Right? An identity fabric where we're integrating key IM capabilities to build a strong ecosystem.

So, when we look at AI, a lot of organizations I talk to are looking at AI as a front end piece to IGA or PAM. But we can actually use that as an embedded component to an identity fabric from multiple areas. Give you an example. I need access to the customer analytics dashboard for the Q4 report I'm working on.

Well, what can AI do? First of all, let's check their current role and access. It can go off and query all the current roles, similar requests, sense of classification, see if it needs privilege access, build a complete business justification on the fly, generate its own risk assessment, determine if MFA is required, and if so, go ahead and do that. Verify a reporting period, and then when it's done, grant or deny that privilege instantaneously.

So, it can actually intersect and integrate with all those components inside of that identity fabric. Pretty cool. All right. The third real world case. Top five most reliable car manufacturers globally. Top five. Number one, Lexus. Number two, Toyota. This is actual, this is true, by the way.

Subaru, number three. Four, Honda.

Five, Mazda. What do those all have in common? Japanese cars. Was that a coincidence? I don't know. But what I do know is the Japanese business model uses a principle called the Kaizen method. What that means is we're going to continue to make small, consistent changes and significantly improve what we do over time, right?

So, if they build engine mounts to hold an engine in a car, and they figured out a way where it can hold 1,000 kilos, well, what if we can engineer that to hold 1,200 kilos and then 1,500 kilos? So, constantly making small improvements. That is how we should look at AI as we're actually building out our modern infrastructures with IAM.

So, here's something I did. This is an actual quote from me. Create a diagram layering AI over an identity fabric, and I wanted it to include things like ITER, governance, et cetera. The first version it created for me was that.

Now, when any of you guys use that and present that to your boss, my boss is in here somewhere. I don't think he would like that. But when I got and applied my Kaizen principle, I then said, okay, include agents for AI, human governance layer, core services, et cetera. Added a whole bunch more, continually trying to improve it, and then I got something like that.

So, this is something that you can always remember that AI has the capability of is constantly making improvements over time. So, applying that Kaizen method.

So, in summary, really, we don't want AI to replace deterministic IAM policy. Should we ever allow it to make runtime decisions? Probably not yet, right?

So, remember, the security controls have to be predictable. It's going to enhance your IAM platform, but it should not be the foundation, right? Should not be the foundation.

So, let's come to our three points. Hyperautomation tool. Look at it as hyperautomation. Don't rely on it to build and maintain. Deploy faster. Test your policy. Thought generation.

Remember, seven days to seven minutes. Strategically, don't look at it as just a front end for individual platforms. Inject it into your identity fabric from all areas, and be prescriptive. Use as much detail as possible, continually trying to improve. Patrick's out here to kick me off the stage. Thank you guys very much.