AI for IAM in practice [Introductory]

Thank you, thank you very much. Well, actually my role is Inter-IKEA IAM part lead at IKEA and we do flat packages as well as round meatballs and emotional support animals across the globe.

So, how many, let's see if we can get the mic to work. How many are tired of AI? And you still want to, and you still want to talk, it might be the microphones here, so I move a bit, that helps. And you still want to talk on AI, something in you wants to listen more on AI.

So, when we talk about AI in an IAM context, we really have two different parts. So, the first part is AI for IAM, which this talk is going to be about. And the second part is IAM for AI, which Jungel Skog is illustrating here.

So, most companies today are implementing AI in some form. Either the leadership has embraced it or part of the organization has embraced it and are happily implementing it anyway. And what this talk is about is to say, okay, in theory, when you roll out AI in general, you are, of course, hopefully getting some budget to do this. But not always, especially in financially tough times, as well as you're probably not going to get all of your budget that you would like to have.

So, can we leverage AI in our IAM operations to get higher productivity and release the resources so we can take care of Jungel Skog here in IAM for AI? Because if we don't manage AI in generality, or if we try to manage AI like we used to manage humans, we are probably going to get in a lot of challenges. Because we have always had problems with managing just normal non-human identities. And most likely, AI is going to make that problem at least a factor 10 bigger.

So, what I'm going to do in this talk, I'm going to talk a bit about what we have used and seen working, and also what our peers are seeing working. So, we're going to talk about identity security, ITDR, natural language interfaces, application onboarding, and configuration optimizations.

So, these are some areas that we are using today that are working well, where I will speak to my peers, and I've got good experience. This is by no means a full set of things you could do with AI, but these are things that we have actual experience of.

So, let's start with identity security. And this is something that we have spent a large amount of our time over the last two years. And the reason for this is that most attackers, the reason they attack is that they want to turn a profit.

Of course, if you're in a business that has to do with national security, you may see nation-state attackers are similar. They don't really care about profits. But for 95% of all companies, a lot of the attack is going to come from people who wants to make a profit. And the most cost-effective way to break into most enterprises is to take over identities. That is partly because in most companies today, the identities is the control plane.

So, once you have an identity, you can usually get quite a lot of access. The other part is that it's usually the most cost-effective. Getting a day zero is not that easy and not that cheap.

Still, at least, we'll see what happens once the AI comes out and start looking for day zeros. But I think today at least, and most likely tomorrow, identity-based attack are going to be very common. And AI is frequently used as part of identity attack.

So, most attackers today uses AI to craft the phishing letters. So, you get very nice phishing letters with good grammar, good spelling. You could even get the right tone of voice.

So, at IKEA, we have an IKEA tone of voice. So, we kind of train how to write things in a special way. And you kind of, after a while, recognize that. And you could get phishing letters that are written with IKEA tone of voice.

And, of course, the old saying that, you know, phishing letters often have bad grammar and so on. That's no longer true if it has bad grammar, bad spelling. Most likely from HR, so it's totally fine if you click on it.

So, how do we do this? How do we defend against these types of attacks?

Well, of course, you still do the training of your people. And you still do, for example, e-mail scanners that removes e-mails.

So, for example, my e-mail that I had for registering this presentation for a speaker. It lost its link to the speaking registration page after a while. I discovered that, like, it took about five to six hours. And then that disappeared.

So, it clearly works. But what you do is that you say, okay, the attacker gets your identity. But can we look at other things?

So, if you log in using a corporate device that says that you are a managed device that is compliant with company policies. You log in from an IP address that comes out of the egress IP out of your network. You use the applications you usually use. And there's no indication on a dark web that your account has been compromised. That's probably quite low risk. On the other hand, if you come from a place where we don't do business, brand new device, you're a frontline worker that suddenly has developed an interest in logging on to the treasury systems.

And you do it in a time that you don't normally do. And we're seeing that your account has been, is for sale on a dark web.

Well, then we probably want you to do something. Like, for example, the self-service password reset.

So, we try to then put steps that are expensive for the attacker, but cheap for us. So, the end user needs to spend 10, 20 minutes to access the password reset. Then all the tokens get invalidated. The attacker loses the token.

Of course, assuming that you have set up so that you cannot register MFA devices from anywhere in the world. Because then the attacker is evil and registered their own device. And then they just reset the password again. This is an excellent example where AI works really well.

So, you can fully automate this. And, of course, you could automate it with rule-based as well.

But here, the particularities of AI is actually quite good. Because you can build some kind of point system or something like that with rule-based. But AI can kind of figure out what is most important when in a quite efficient manner. Another problem we have in the IDA system is that usually when you roll out an IDA system, you might have 20, 30, 50 applications. And then you usually run out of money or more applications.

So, either money or energy. So, that means that when someone leaves, you can, of course, remove all the access that they have on the person. And we can take away their main IDP or main IDPs. That usually takes out with SSO quite a lot. But you most likely have a certain amount of dark applications that you don't know about.

So, then you have a system where you tell, hey, GCP team, please remove the access. And this, of course, is also more complicated that you have personal access. And then you have NHI that you control that needs to be removed. And then you have agents on top of that that also needs to be removed.

So, if you have a situation where you need to walk someone out the door and need to do this quickly, and also generally just doing it efficiently, being able to get the overview on IVIP potentially using AI for doing these searches, that is quite useful. And, of course, if you need to do this, but in a situation where you get a call at 2 a.m. because you have a compromised senior engineer account, it's a little bit like going to Ikea to Smaland and go to the ball pit. And then you say, oh, there's 50,000 balls. Fifty of those are slightly different color or slightly heavier.

And they are the things you need to find. It is not a fun thing to do at 3 a.m. in the morning when the CISO is sitting there and going like, yes, you need to shut down all access to this identity.

So, again, here is where AI can really help. And, yes, you may not get everything, but sometimes getting 95 percent or 98 percent is a lot better than 10, which is what you usually can do manually. Natural language interfaces.

So, I'm quite old. So, I started out doing things in natural language interfaces in the 80s. And it is interesting that you're kind of back to this type of configuration.

So, normally we configure our IAM system with either code in some form or click-ups. What we are seeing now is that the natural language interfaces are really coming. And this is really useful, especially for things like reporting.

Like, IDA systems are notorious for being bad at getting their data out of them. So, usually I have to write SQL code.

So, being able to just ask a question and then, like, a minute later you get a report on something instead of having to spend, you know, a couple of hours with your favorite SQL client or LDAP client getting the data out of different systems. Or shipping something offshore and getting it back a few days later. That is really powerful. Most modern IDA system either has natural language interface today or are getting it. If you have a bit more of an old school system, there are specialized vendors that does things like this, like IAM1s, for example, that you can put on top of your systems.

And hopefully, you will not have the problem that your natural language interface gets sent to them to stop answering your questions when you start asking evil questions. So, that's how the IDA install you. Another part of similar things is configuration optimization.

So, any system that is sufficiently complex is indistinguishable from magic. And when you come to things like a mature, large enterprises configuration system like Microsoft CA, for example, you kind of are getting into magic territory.

So, by being able to run optimizations against this, and historically how we did this is that you get to do a zero trust assessment or equivalent. Those take quite a lot of time and also resources. And if you go out and buy them commercially, they are quite expensive. What you can do is to connect your systems to an MCP environment. One of the first one specifically for Entra Conditional Access is Locka. And then you can start asking questions. And we have been doing this on some of the test environment that we have that contains no playgrounds, sandboxes. And we got some quite good outputs.

Even just having a actually in basic copilot is quite good at that. We used as a super powered search engine. You can get very good answers from that easily. One thing to be careful about is, of course, if you connect it to any LLM on the market, even if you have it on read-only, the problem is that your configuration just ended up in the learning system, this LLM. Which means that your configuration information is now happily swimming out there in the general internet.

So, not a good idea. So, try to get an LLM that is corporate and that we have clarity around where the learning information goes to. Challenges.

Well, the first challenge is non-determinism. So, normally, you know, the definition of insanity is that you do the same thing again and again and expect different answers. When it's AI, it's actually how it works.

So, you will have non-determinism. And the question is there, can you live with that?

So, if you look at the first example on identity security, you don't, in practice, it doesn't really matter because it's such a complex system. So, you get non-determinism practice always.

But, of course, you don't want the system to be able to make decisions that actually matters and then having to explain that I don't really know why it did what it did. Prompt engineering. It takes a little bit of training to know how to do this.

So, do give your teams a little bit of time, yourself, to learn how to write prompts well. So, that's also a skill that you have to learn. Boundary definition.

So, this is an interesting thing because this is actually a quite hard thing. So, I have my own personal AI as well. It's also named Snuttig. And I wanted to go to, I used to be a stage manager in my previous life. And I wanted to find a club that's played post-punk and goth music in the Amsterdam area.

So, I thought, hey, I can ask my AI to find this for me because there aren't that many of them. So, I did that. And I had seen in my Facebook feed that there are some clubs every now and then.

So, there should be something, you know, somewhere. So, I asked that. And he said, nope, nothing. Okay. But I did see that in my Facebook feed.

So, I went back and checked. And it was one in Haarlem, which is just outside of Amsterdam.

So, I asked Snuttig about this. And he said that, well, he didn't really feel that Haarlem was in the Amsterdam area.

Which, okay. But that's something to think about. He was very, very good at putting himself in this goth dress.

So, it's all about, you know, asking the right things that they enjoy doing. Outdoors and source checking. This is also an important part. And you can see that you get kind of seduced by the easiness of just asking questions and getting responses. But AI tends to hallucinate. If it can't find an answer, it can make something up.

And, of course, if you ask for something where you say, okay, can you please get me the standard for something? Like for password management according to a specific system. And you can, of course, check the sources and see that actually what it's talking about is correct. But this can be much harder.

I, for example, tried to have my kids lost all the remote controls from my garage door. And I got the garage door with the garage when I bought it.

So, Snuttig here thought that, yes, this is a standard garage door. You just need to go and buy yourself a standard remote and then slave it in. And it's going to work great. I spent many hours on this and completely failed. Then I got the professionals in and they said that they had no idea where this come from. It must be some Chinese import. I've never seen it before.

So, something to think about and try to figure out, can I check the outputs? If I don't, if I'm not able to do a double check on outputs, is this important enough that I can accept that? Application onboarding.

So, as I said before, this is one of the things that is very expensive. And it's something that has also really not become cheaper in the 25 years or 20 years I've been doing this.

So, the idea is that you use the IDAs to handle a lot of the process. And you can either kind of automate the questions part or you can even put a little agent on the system and let it sit and listen. And then automatically generate the configuration. And then you can check that with the owner of the system that they're okay with what the AI thinks. Definitely really interesting if you're going to do an onboarding factory. And also very useful for going out and looking for systems if you want to do it in an IVIP configuration.

As you do see here in this generated picture, the onboarding here is not so successful because the applications are jumping into the sea rather than to the ship. So, again, quality control is important.

AI, don't trust AI. Regulations and compliance are coming just like the winter. I'm originally Swedish. And in Sweden, we always feel like it's midsummer, which is in about a month or six weeks. And we always say that now it's getting darker again.

So, just like winter, regulations are coming. The EIA Act is in effect.

Now, it mostly targets things that are really important for the individual's rights and freedoms. So, most likely what we do in IM is not going to be hit first by this. But really good idea. Figure out where your company's AI registry is. Register your AI and put in the owner and intent.

So, what do I do first? Well, I think the first thing is to think a little bit about how can I leverage AI for improving my team's productivity. Let's start experimenting with this. And talk to your peers. Figure out what they have done and what worked for them. The second part is to try to talk to the business.

So, for example, an onboarding factory is really only useful if you're going to onboard a lot of applications. So, try to keep a little bit of an ear out.

And, of course, it is very possible that the business is rolling out a major AI investment. So, try to see if you can kind of, you know, ride that wave to get your investment in place as well. And once you have a good understanding for things where you can improve your productivity, you have an understanding for where the business is going to take AI, you should take a look at your target architecture and your roadmap and see how can I be effective in an AI world. Thank you. I think we have about a minute for questions.

Well, thank you very much, Martin. That really gave me a lot of ideas to definitely try after the conference. I hope the same worked for our viewers as well.

So, do we have any questions? Just a quick reminder. Raise your hand and I will give you the microphone. Just one quick question. I think the element of optimizing or improving your configuration for IAM infrastructure is pretty interesting and very valuable because there's a lot of loopholes that might actually exist. But there's this tension around, you know, public knowledge about a product and then corporate knowledge about what is the setup I had in place or I have in place.

So, just like you say, fitting that information into an NLM or into a piece of AI that is not approved by your corporate organization is involving risks. So, how did you resolve that tension between leveraging AI to improve the setup but at the same time not disclosing what could be actually weaponized against you guys, right?

So, yeah, that's the question. Yes, so basically my recommendation is that we are trying to get a standardized platform we can use.

So, we have today, we're mostly a Microsoft shop, so we have a co-pilot. And Microsoft is handing out a lot of the more advanced licenses where you can also connect an MCP to them.

So, that handout is going to happen now during May, June. So, that is kind of our goal of this. But it is important. I know another company where the outsourced network team decided that they wanted to improve the configuration.

So, they've had the entire network configuration into OpenAI. This was not so happy.

Great, thank you. Okay, awesome.

Well, thank you much, Martin. And in the interest of sticking to our schedule, let's directly proceed to our next presentation.