Yes, thank you very much for the introduction and hello everybody. Yes, today I'm talking about Non-Human-Identities. I agree to talk about this subject. It was one year ago when last time when I visited the conference and since then things moved quite fast. So I ended up, well, it's more or less a lot of, less Backup-Jobs and Non-Human-Identities and more AI-Agents, we'll see.
Okay, so first of all, who am I? For people who don't know me, my name is Eleni Richter. I've got a degree at University of Karlsruhe as Wirtschaftsengineer. It's a degree in engineering, so I'm not native IT. As you can see, I'm the typical German engineer.
Yeah, I'm more than 25 years now in different positions at ENBW. ENBW is a large German energy provider with also some international parts. I've been working as security manager, IT consultant, system designer, project manager and architect. And right now I'm chief architect of the identity and organizational data management of EW and ENBW. And I'm responsible for design of our identity management system. And I also have a part-time lecture in identity and access management and cybersecurity at University of Luzern.
So, what I'm talking about? I'm an engineer, so I like to draw. I like to draw plans for construction.
And, well, I brought my sketchbook. And we're going to have a look at practical approach to non-human identities. I'll just give you a warning. I don't have all the answers you're expecting. There are still some very open ends regarding AI, but we'll see. We'll go through a practical definition of terms. We'll take a look at actors and use cases, agentic AI hype and practical effects.
And, well, we'll look at non-human identities as cybersecurity threats and agentic AI as fire accelerator. And we're also looking at some building blocks and missing parts.
As usual, the disclaimer, everything is fictional. And now we'll start with a sketch number one, a practical definition of terms. Another identity, of course, everyone knows is a set of account and credential used by an actor for interaction with an IT application.
That is, well, very common since I think since this conference exists and even before. Usually we have started with looking at the actor as a human identity, then the actor as a person who is using IT, like interaction with an application. And this actor, the human being, is capable of any kind of complex interaction you can think of, of course. By yourself, you're all human and, you know, you can mess up your IT. So in contrast, we have the non-human identities.
Still, identity means a set of account and credentials for use in IT applications. But this time it is not used by a human, but a machine. I'll start with a simple non-human identities. These are whatever you call them. You know them since 30 years and 40 years.
It's task, jobs, daemon, program, script, service, whatever. They are manually managed and they run automated simple tasks. And somehow they have a config where you place user ID and password and the job is using it to do some automated task in an application. If you're better, you don't place it in a configuration, but you place it in, let's say, in a pump system or whatever, or password vault. There are a lot of tools around. So if we get a complicated non-human identity, that makes it more different. Then we have a job factory which creates jobs on block. And these jobs do perform tasks.
These things started occurring when we went for microservices, for virtualization, for cloud, jobs on demand, or everything is code. We did not create the exact job as code and said, OK, this is a job. I'll place the identity on this one. I want it to run the backup job. I have only one and it does the backup.
No, but we're going for, let's say, microservices in the cloud. We'll scale up on demand. And we got a job factory or a job engine which creates jobs. And therefore, we need automatically managed identities to run automated multiple but still simple tasks.
And, well, that comes up to the fact that we have some kind of config where we store the user and the credentials which should be used for these automatically created jobs. Now, if we go for, it gets even a little bit more difficult. We go for complex non-human identities.
So first, let me introduce something else. AI, that means artificial intelligence. It's a class of algorithms which are probabilistic, uncertain behavior, often open, and in some, it gives us a complex system. Gen AI means simulation of human behavior in voice, speech, language, or video and picture, or any kind of text generation. And text is including code, so the programming code you're using. And programming code, producing code which is programmed and can be executed, means we're producing interaction or possibilities of interaction.
So an agentic AI is an AI with partly autonomous behavior. So that brings me to my definition for complex non-human identities. Complex human identities are AI-based actors. They are complex actors. That means they are a computer program using AI for complex interactions. That could be complex interactions with a classical application. That could be complex interactions with other complex actors. So your favorite chatbot or whatever. But it calls also other complex actors could also, again, be an artificial intelligent agent.
Okay, now we got three types of non-human identities. If we go for actors and use cases, as I said, inside here we are placing some IT system you want to access, you want to analyze, you want to have interaction with. And if we just go around, we could have good actors. We distinguish good and bad actors and interactions. So everything which is black is somehow good, but if it's contaminated by some red color, we regard it as a bad interaction. That could be a hacker who is accessing, trying to access this IT system.
This could be a botnet, part of a botnet, malicious botnet, trying to access, to do some malicious things inside our IT system. But it could also be a good guy here who is doing something he really didn't want to intend to do and is doing something bad. So if we're going for a very plain and simple security model, it's about regarding a cyber system, allow only good actors doing good actions under wanted environmental conditions. So it stays up here with the green ones.
Okay, if we've got a human identity, a person, we've got a human background. So if you're just looking around, you see a lot of examples sitting next to you, humans with human background. Human background means you have emotions, you know fear, you have hope, you have an education, you have experience, you have something like a common sense, you have moral standards, ethics, you're embedded somehow in a legal system, which was created for humans, by the way, and you maybe are bound to some contracts or whatever.
And, well, this is something which I would sum up as a human behavior, and it's somehow something like 2,000 years of civilization or something around here built on. And then we've got some well-known risks which occur in IT environments. And let's say, let's go for one of the OWASP top ten lists. And that is why we've got a, if we have got a human interaction, we've got a behavior management by civilization. And we've got some catalogue-based patterns for enhanced security like the OWASP top ten we put on top.
But the main thing why you know how to behave is not because you've got some security catalogue from an OWASP list or whatever, because you've got an education, you're part of the civilized world and you're a human being. So if we go for the non-human identities, well, we've got this, in part A, we've got this simple job. It's based on a deterministic algorithm. A deterministic algorithm is an algorithm which says, okay, to get from here, from point A to point B, that's the way it's described and it runs always that way. And it's a deterministic behavior, it's predictable, it's static.
And, well, there are well-known risks. They can be easily described or they can be described. Let's say you can go for the OWASP top ten non-human identities. There we will find a very good described list of what could happen. So if one of these actors would turn bad.
Okay, and we've got a simple tooling, for example, secret vaults or inventories or something like that. There is tooling in place, it's nice to have. If you go for the complicated non-human identities, we've got this job factory. It's still based on deterministic algorithms. We have some adaptive parts. And it's complicated, but it's mostly deterministic behavior. We have still well-known risks. This might scale up a little bit in the number, but it's still well-known and describable. And we have enhanced tools for automation, like managed identities.
We can put them in the pipelines or whatever. So that means for part A and B, we've got security management by catalog patterns, which describe the do's and the don'ts. We should do or we should not do. And we have some nice and powerful tools in place. For example, the speaker before me just showed you one. So where's the problem? What about, see, I left them out. The complex non-human identities.
Yeah, now it's getting complex. So that's our landscape we have been looking for the last 20 years, and it's nice. But what's now happening? The agentic AI hype and practical effects. So we've got this complex actor. And it has the background of this special non-human identity. It's based on AI algorithms. They are non-deterministic simulations of humans. And non-human...
Yes, they are non-deterministic. They are simulating human behavior. They do non-human interactions. And they're incredible flexible. And they have a dynamic behavior. So if you put them in this use case model, that means, well, how did this complex actor has his knowledge on how it should act? How did this underlying model was trained to work?
Well, the underlying model was trained to work since we got all the information we have in internet and we used it for training data. We put it into the brain of this complex actor. So we have data on any kind of actor and interactions inside this algorithm. That means we don't have only good interaction and good actors inside, but we have also the knowledge of bad interaction, of stupid interaction, of just unwanted interaction. It is somehow inside here. So the behavior management by civilization...
I mean, you have it also in your brain. You know things you should do and you shouldn't do and so on. But you have it in your brain. But you have some behavior managed by civilization. But this is not civilization. This thing didn't get an education or something like that. And whenever you train him on the data, it turns somehow that it has some of this bad knowledge inside. Or it may put things together which are good and the outcome is bad. So if we have use cases with complex non-human identities, we have complex interactions with applications.
And remember, the color red is for bad or unwanted interaction. So this thing is just capable of doing bad or unwanted interaction. Not only with applications, but also with complex actors like other human beings or other agentic AIs. Could have another agent like this. That is just the picture of what happens if a complex actor is doing complex action sets.
And now, if we go for our security patterns, the patterns we used to use in situations, we go for this fixed catalogue patterns for security. Any OWASP list or whatever. Can we really address this incredibly flexible thing with fixed lists? And if that thing is really simulating behavior, we'll need behavior control. So non-human identities are a cyber threat. And I consider the AI as a fire accelerator. And with simple static approaches, we won't control these complex action sets. But what should we do? We'll take a look at some useful building blocks. We take again our complex actor.
But as we've seen, not useful are simple static approaches which cannot control complex action sets. Choose wisely where you use your limited resources. So necessary, but not sufficient. Any catalogue of patterns or anti-patterns, good or bad ideas for security, also apply on this complex actor, like an OWASP top 10 or whatever, OWASP for agentic AI or LLM. Always a good idea. Have a good lifecycle management for complex actors. And maybe you also want to go into some off-record usages and go for hardening your application against threats which come derived from a complex actor.
But that is not enough. Useful but limited. If you go for OpenID Connect, MCP, OBO flows, and a very strict identity management, all the protocols are just started to integrate the agentic AI things, and that's useful. And a very fine-grained access management is also a good idea. Fine-grained authorization, human-in-the-loop, might be a good idea. Does anybody remember this famous authentication fatigue attack?
I mean, you just send a person who is using an authentication of these push messages, authenticate now, authenticate now, authenticate now, and the person doesn't control it anymore, but it presses just, I will authenticate. So I do not believe too much in human-in-the-loop because we can't control complex actors and complex actions by hand. It's not really. So send-back boxing for the identity. Use disposable, non-transferable, non-Hume identities. But control and automate by clicking. That is the thing with human-in-the-loop that is very limited.
If you're going for disposable, non-transferable, non-Hume identities with a very short lifetime, and I mean really a question of minutes, control something fully automated by a short lifetime, really? I mean, two minutes are endless for a computer. It can do many things in that time. It might be very short for a human being, but it's really far. That won't work too much.
And we shouldn't forget the whole is greater than the sum of its part, and if this actor is doing one action, another action, another atomic action, that the whole could be a real problem, and we can't address it with fine-grained authorization because we then get only the bits, and we say, okay, it's okay if reading there, it's okay reading there, it's okay writing there, but we don't know what these three actions in sum would give, and maybe they would cause a problem. So to sum it up, humans tire easily. I'm flexible. I will find a book around, because it's not a static algorithm here.
It's a flexible thing. Humans can't imagine what I can do. I perfectly simulate humans, so if you want to do something with behavior detection, well, that thing can simulate human beings, and I know camouflage and deception. This algorithm has all this knowledge in its brain. So useful approaches are, of course, zero trust, and if you go for, if you want to analyze a situation or where you want to use an agentic AI algorithm to do a job, you may want to go for a system analysis because it's a more flexible way in dealing things.
You will go for resilience, which is what I could go for a whole contract, and you want to invest in the upper-level security application layer and business process layer. So you should enter control goals, scope context, sandboxing, but there's still a problem, and missing future aspect, we need sufficient security controls and business and application layer. We'll need AI-based mechanism to govern the behavior of AI algorithms, and in the end of the day, we'll need some kind of HFA.
I called it a human factor authentication, so we got an authentication proof of bio uses and no fake identities. We have to wrap up. So recommended reading. I put in some papers if you want to continue, and final word. Yep. There's still a lot of things to be done, and if you have any questions... Thank you very much. We don't have any questions here, but I assume you can track her down after.
However, given you're a professor, I've always wanted to do this. I assign you for next year. Come back and tell us how to manage these complex energize.
Okay, thank you. Thanks very much. Is Daniel here? Thank you.
