Welcome to the KuppingerCole Analyst Chat. I'm your host. My name is Matthias Reinwarth. I'm analyst and advisor with KuppingerCole Analysts. My guest today is Alexei Balaganski. He is a lead analyst and the CTO of KuppingerCole Analysts.
Hi, Alexei. Good to have you.
Hello, Matthias. Great to be back again. Thanks for having me. Great to have you. And as always, we will have a great topic. And this time, first of all, the main topic is we want to talk about AI agents in cybersecurity, the role that they can take over. Is there a revolution dawning? And don't go away. By the end of this episode, we will have maybe an answer to the question whether there is a magical or tactical solution for the cybersecurity skills gap. So let's find out whether we can get there via the AI agents.
But first of all, how do you, before we start and digging deeper into the repercussions of this topic, how do you define AI agents? Well, I guess, Matthias, we should step even further back and kind of explain again in a few words, like why are we even talking about the agents or like, why is everybody talking about them nowadays?
Well, first of all, because we finally have them, like there are already working solutions offered in various fields of applications, not just in cybersecurity, of course. And finally, people are getting this feeling that it is actually something big. Some are afraid, obviously, because they fear that agents will just push them out of their jobs. The others rejoice because they will cut their costs significantly. And of course, the vendors are happy to push you their new products, new buzzwords.
But what we want to talk about today is a kind of a more philosophical or strategic way to see where does it all go in the future. Is it really the next revolution, not just in cybersecurity, but in the whole IT industry or like the whole industry period, just like the cloud before or now the steam engines even earlier on that.
But yes, you're right. AI agents, we have to define what they are. And by the way, we have to differentiate agentic AI as a scientific academic concept and AI agents, which are actually tangible solutions, which may or may not necessarily even implement the concept of that agentic AI. But basically, a typical AI agent is just a piece of software, it's a program, but it's substantially different from traditional applications in a way that it can not just follow a predefined set of rules, but it can basically make up its own rules on the way it can perceive the environment it works in.
It can reason about it, make its own decisions, again, not based on any predefined rules, but on the dynamically emerging logic. And finally, it can take actions. And of course, those actions are the biggest thing everybody is now talking about.
MCP is, of course, the next buzzword, model context protocol, the way we can allow any AI agent to interact with the environment, to connect to a database, to send an email, to shut down another application. Basically, we are almost at that level already where the agents basically walk into your company like interns, they take over some of your easier jobs, and you have to be always vigilant and take care that they don't break anything. This is where we are at the moment.
These are pieces of software that behave differently, that have a sense of the environment, that understand what they are expected to do. They have goals. Maybe they set their own goals or sub-goals. That would be the main characteristic. It's no longer just an if-then-else thing, but really a more agency-based, that's where agentic comes from, agency. They are acting on behalf of themselves, and then on behalf of somebody who has given them an instruction. Where do you, you are the cybersecurity expert, where do you expect these AI agents to actually be useful?
What are applications where they immediately or in the foreseeable future can provide benefit? Well, first of all, again, we have to reiterate that the biggest difference between an AI agent and a software agent, if you will, is that they act like humans. They have at least some symbols of a brain, thanks to a connected LLM, and they have their own actuators. They have tools which they can directly use to manipulate the environment. So obviously, if not necessarily today, but some hopefully not very distant future, in theory, an AI agent can do anything.
And there are some applications where they are already good enough. So obviously, one of the things everybody is talking for the last at least five or even maybe 10 years is the skills gap. There is not enough humans in the cybersecurity industry to man every security operation center, to respond to every alert, to fix every data breach, and so on, at least in time. That's kind of the biggest challenge. There are too many things happening and too few people to respond to all those challenges in time. And this is where some vendors are already bringing in AI agents as a promising replacement.
So obviously, if you have a security operation center, you know that they usually generate tons of alerts. And we have already lived through several generations of tools, the next-gen themes, the source, the UEBAs and whatnot, whose only promise was basically, we will reduce the number of alerts you have to investigate, because the rest is hopefully just noise.
Well, AI agents give you an alternative, where you only have like five human analysts now, you can have five AI agents. They would work probably a thousand times faster, and they would cost hundreds of times less than a human analyst. And if five AI agents are not enough, you can have 5,000, thanks to the power of the cloud. So in theory, if you believe the forward-looking statements of some vendors, there will no be the skills gap as a problem anymore, because anywhere we don't have enough humans, you can just employ AI agents. Are they there already? Probably not.
But I've already seen very interesting applications where an AI agent has been successfully taking over a very specific, like one specific type of investigation, but it does it so well and so quickly that you no longer need any pre-processing steps, no longer need to filter out all the noise, because during the time where a human could investigate one alert, an AI agent can do a thousand. So it would happily take over the entirety of your, let's say, email attempts and verify every suspicious email, instead of looking or waiting for some other tool to filter the noise out.
So this is where you already see AI agents really kind of taking over an analyst's job directly without any modifications. They would just use the same interfaces, or they would just connect through an MCP standard interface and directly replace a human in an already existing piece of software. No major changes needed. That's kind of their biggest promise. Right. And as you've mentioned, we've made podcast episodes earlier around exactly this topic of drilling down, of boiling down the full number of events that come into those who are important.
And I think when we do not have to do this analysis process, is this important or not, which comes with the threat of you miss something or your filter is wrong for boiling down, you now say, okay, no, let's let the machine, let's let the agent check all of them and let the machine learn what the actual patterns are. And we don't have to superimpose that in the beginning to say, yeah, I don't know this. I don't think we should investigate that. And maybe that's an error. So it's getting better and at a much higher volume, right?
Well, you know how they say that even an infinite number of monkeys with typewriters would never recreate Shakespeare's work. But there is a lot of cybersecurity jobs where you don't need a Shakespeare. You need a person who can basically read basic rules and follow them. And those AI monkeys are already good enough to take over and they are already doing that.
And again, thanks to the scale of those cloud based, centrally managed AI agentic solutions, they can grow, they can meet any kind of demand, whether you are a tiny small business or a huge international corporation with millions of alerts daily, they can grow up, meet the challenge as long as you have very specific, easy, narrow tasks. But of course, with the time you would expect them to become more sophisticated, more generalized or simply to have vendors who would offer you instead of five different types of agents, maybe 500 different types.
Again, if sometime in the future we would evolve, that whole technology will evolve to create and train an AI agent for a specific task automatically without any human involvement at that stage, then again, the system would scale automatically in both directions, both horizontally and vertically. So the possibilities are endless.
Of course, we are yet to see how well it works in reality, but the promises are really revolutionary. This is why I would argue you really have to consider it as a possible next industrial revolution, at least in the IT industry. Right. And boldly ignoring the ecological impact of having lots of these machines running all the time. But if we go back to my original questions, if we have those 5000 properly trained, virtualized monkeys that are really good at individual jobs, what would be the jobs where they are good at right now?
What are applications where the end user, our audience, can expect that there will be at least an augmentation of support by these monkeys in this work? Well, one use case you already mentioned, it's basically automating incident investigation and response, although things which are now usually called threat hunting and triaging and digging through logs and what not, following the playbooks.
And yes, some of those playbooks are already automated with like SOAR solutions, for example. But those playbooks are usually very rigid. They will break every time something unexpected happens or simply when your environment changes. Whereas AI agents can naturally adapt to those changes because they are already flexible enough and they do not need to be retrained, they do not need to be reconfigured. Because again, even the process of figuring out which tools they have and which inputs they can consume is already handled by them without any human in the loop.
This, by the way, kind of leads to the biggest question yet. Do we still need a human in the loop in security incident investigation? This is not a technical question. This is still largely an almost religious decision and a lot of companies are still fearing that kind of autonomous activities by the agents. But sooner or later, they will have to get used to that because, again, the scale of modern IT is still growing exponentially and we already don't have nearly enough humans to be in the loop. And of course, there are many other use cases.
Vulnerability management, proactive maintenance, hardening, posture management, basically all that analysis, which now is usually handled with some kind of schedule-based tools, which usually only work, let's say, like daily because there is not enough power to do that every minute or in real time. Again, with AI agents and the additional flexibility and dynamic ability to respond to changes and unexpected new factors, they will scale much better than traditional tools. The same goes to identity and access security.
This is probably the most interesting topic for our listeners because that's what we usually cover. But yes, all this threat intelligence and identity, threat detection and response, and just kind of making sure that your existing policy-based access management tools actually work properly, that the policies are not violated, that there are not just suspicious anomalies to investigate, but you can actually investigate every possible transaction, hopefully in real time, but even if not, fast enough to actually prevent a potential breach before it happens.
Because what's the point of doing threat hunting and finding out that, yes, you were breached two weeks ago, if you could instead identify it within 60 seconds, for example. The potential damage, the risk explosion radius will be so much reduced that could be a big difference between a compliance fine or a full blackout in your digital business as opposed to just, well, stumbling but standing up and going forward. Maybe that is also, which is almost paradoxical to say, okay, we are still looking for a solution for how to manage access for agentic AI, for AI agents.
They behave, as you said, they behave like people. They have the same behavioral patterns that they have. But we do not know why they behave, how they behave. Having to send a thief to catch a thief, so to send an AI to control the AI might also be a first step or maybe the main, most important step to also apply identity and access security to those agents that we're just talking about.
Well, this is exactly the same kind of mentality a lot of people still share, like, but what if we have a Skynet scenario, like in the Terminator movie, but what, like, who watches the Watchmen? Like, how do we know that the AI won't go rogue and hallucinate and break something?
Yes, they can. They absolutely can. And they will. The problem is they are no different from humans in that regard. The only difference is scale. A human can go rogue, a human can hallucinate. We have done that many times in the past.
I mean, we are humans after all. The whole point of this development and all these discussions is that we have to design our entire access management and security and observability and other infrastructures to be resilient to both humans and non-human agents.
And again, yes, this is a scale where just a human alone cannot keep up. So it has to be at least performed by another AI agent. But a human should be in the loop in the sense that it has to be on the higher level of abstraction and visibility. And instead of basically validating every activity of an AI agent, a human has to validate the strategic development, that the goal of the specific agent still aligns with your business goal. That agent is not allowed to do things any other agent is not supposed to, or any other human, if you will.
Like if, for example, you are using an AI agent to test your business critical application, the agent should not have an opportunity to drop your production database. But the same applies to all the human administrators. If a human administrator cannot drop your database, but an AI agent can, then obviously something's wrong with your access management, not with your AIs.
And this kind of brings back this whole, I guess, idea of you have to have a fabric, you have to have a strategically designed infrastructure, which puts all these components together and make sure that they not just perform properly, that they have optimal coverage, that they align with all the policies, and that the policies are correct and consistent, but it all works the same, regardless of scale or nature of your agents. Is it a human? Is it a script? Is it a cloud service? Is it an LLM on the AI agent, or is it 50,000 AI agents?
In theory, it should not matter. Somehow we have to design this whole thing, that it works the same all the time. Do we still need a human to do that? Hopefully. But can a human do it without AI? Probably not. So we have to go on a higher conceptual level and start thinking in terms of, again, governance, I guess. How do you govern your AI developments? How do you govern your AI policies? How do you govern your governance, if you will? How do you govern your compliance, if you will?
This is as much a philosophical discussion as a technical one, but there is so much to think about and to discuss and develop in the future. It's also about model governance, so we're leaving the cybersecurity aspect as well and move more to expertise in the AI fields as well, because that will influence that as well. But you are talking to vendors, you are watching the market, you hear what vendors say, what other analysts say, what our peers are talking. Are we at the brink of the next revolution in cybersecurity?
Well, if you only analyze what people are saying, you would probably be confused a lot, because obviously different people have wildly different understanding of the whole challenge and wildly different ideas where it should develop in the future. I've heard a lot of companies claiming that they know that they already figured out how to do this, but obviously what they deliver is far from it. It's almost a kind of the next generation snake oil, if you will. Some companies have this figured out for real, but for a very narrow and small range of tasks so far.
As I mentioned, for example, you can already go out and buy an AI agent or 10 or 10,000 to basically replace your tier one security analyst in your SOC. That's already possible, but for a very specific set of tasks, and I imagine for a very specific set of SOC software platforms.
Of course, those vendors will implement more connectors or broader support for the usage scenarios and so on. But this is probably like the most best-sourced use case for cybersecurity. I imagine a lot of automation, a lot of intelligent automation is happening in the area of threat intelligence collection. That's exactly where kind of already traditional generative AI is best at sifting through tons of unstructured data, looking for nuggets of useful information. This is obviously something which is, if not already there, but coming very soon.
We should not forget that the same thing is actually done by our enemies, the hackers, the spies, and just kind of the generally malicious people out there. They have the same tools at their expense. They are already working on better deception tools, infrastealers, scrapers, and so on. And we already see that, we see how modern phishing emails, for example, are so much more convincing.
Phishing, business email compromise, and so on. It's amazing how easy it's now to kind of make sure that a person would be non-deviser and click the link and not even know that they have been hacked. And then again, it goes beyond the security intelligence. But if a user clicks on the phishing link, does it matter if it was a human or an AI or something else? It has to be stopped, or at the very least, nothing bad should happen. And this requires, again, a combination of different tools. You have to have your data encrypted.
You have to have access management to prevent leakage of sensitive data. You have to have monitoring tools. You have to have access policies. You have to have access governance. You have to have threat intelligence. Everything of the existing tools is going away. They're getting better. They're getting bigger. They're getting faster with the help of AI agents. But essentially, yeah, it's basically like an evolution taking an unusually large step. Will this kind of huge step in quantity mean automatically that the quality will increase?
Again, it's difficult to predict. In some areas, definitely. In some others, probably not.
I mean, next year, I imagine, will be the time of really interesting announcements. Right. I really expect that as well. But I really like the idea that you are not distinguishing between humans and non-humans when it comes to autonomously acting actors. So this is really something that needs to be watched and where the same guardrails need to be applied, no matter whether it's an agent or it's a person, because we all can make errors, including those agents. And what I like is also the idea that traditional, in some terms of traditional, platforms won't go away.
There won't be a simple AI replacement for a seam, for a saw. In the near future, at least, there will be this combination, this collaboration of different types of autonomous actors working together, be they human or agentic or agents, to stay with the correct term. When it comes to the platforms that people already have, they have been talking about automation for years right now. But what we are seeing right now is really a different type of automation. So this traditional pattern-based, rule-based automation is just different, right?
Well, remember how we used to have this thing called big data. You would have to have a completely separate database and all the infrastructure built around it for your, let's say, financial transactions, and a totally different hardware and software stack for your analytics. And that analytics would be the big data. And only the richest and largest companies could afford having that kind of thing.
Nowadays, MySQL database can do the same stuff. And it would be like a thousand times cheaper and a thousand times faster than those big data solutions from 20 years ago. I guess the same will happen to other kinds of solutions. So right now, we are still thinking that there are quote-unquote traditional tools and the quote-unquote next generation AI-based tools. Maybe in 10 years or even faster, we will have just tools.
So we will have automation and it would offer you like a full spectrum without any gaps ranging from simple scripting tools on one end and sophisticated multi-agent-based systems on the other end. It's everything in between. Ranging from like low-code, no-code platforms to enterprise-grade ETL, what not, data processing platforms. It would all fall into the same spectrum and it will be all powered by the same kind of hardware and software and AIs, if you will.
I guess if history shows anything, that everything in IT tends to be consolidated and converged and just kind of growing in scale this time. This is what we should expect. So traditional themes and SOCs and SORs, they will not disappear completely. Maybe they will change their names. Maybe they will be acquired and incorporated in new kinds of tools or maybe it will happen the other way around. Maybe the same SOR solution which you love and hate for the last 20 years will now have AI-based and agent-based capabilities included. With a different name, it would still do the same thing.
It will help you to protect your company from attacks and it doesn't matter how exactly it does it. Hopefully, it will offer you a flexible combination, a hybrid approach, if you will. Simple things for simple tasks, complicated solutions for complicated tasks and the freedom of choice to combine them any way you want. Just two weeks ago or three weeks ago, I've been in Munich for the Identity Fabric Impact Day and I did a short presentation. I'm fast in this to say, okay, what we are delivering is a service and a set of services or a portfolio of services.
Then it was IAM, but it's exactly the same for cybersecurity. So we need to understand which are the services that the organization needs, what are the capabilities that are required for implementing that services and what are the tools that I need to implement these capabilities to provide these services and that will change over time. We need to start with problems and we need to figure out what ways of solving those problems we have.
And then you have to basically build the most efficient combination of tools, which without too much overlap and without too much overspending, cover all those problems and capabilities. This is exactly the definition of a fabric, whether it's identity or cybersecurity or automation or data management or anything else. You are thinking in terms of a fabric, which is great. Right. And we will need people to evolve that and provide the right services at the right time for the right group of stakeholders.
I promised at the beginning that we will have a look at the answer to the question, can AI agents be one or the key solution for at least mitigating the skills gap issue? We talked about that in the beginning. From your expectation, will there be less need for seasoned IT security experts?
Well, yes and no. As usual, the real answer is a little bit more complicated. So obviously, there will be some positions, some jobs, which will be completely automated out of existence, just like it happened with basically people moving around stacks of papers. As you know, you just reminded me, I recently attended the Bletchley Park Museum, the place in England where they used to crack the German Enigma machines during the World War II. And at the peak of their operations, they employed over 9000 people just to do that, to crack encrypted messages.
Most of those people were women carrying around stacks of paper, motorcycle drivers to carry those papers back to the admiralty or whatever it was based on. And then they brought in a computer and the computer replaced like 95% of those humans. Exactly the same will happen here. There will be a lot of tier one jobs, which will be completely automated and replaced by agents.
But of course, inevitably, what would happen that you would need a higher level, more sophisticated and of course, probably more expensive human involved in actually organizing and orchestrating and planning and architecting those agent based solutions. If you want to have a guarantee that you will not be completely replaced by AI agents, well, you have to work a little bit on your education. Make sure that you are smarter than any AI agent, even if you will never be as fast or as scalable, but at least you can be smarter. Right. And that will be the shift that you've mentioned before.
There will be new roles required to augment this automation process with new areas such as, as I said, model governance, you said policy design, policy assurance, access governance on a higher level, supported by tools. But in the end, being that human in the loop, that is where the jobs will be. Right. You're talking about new competences, obviously, but we're also talking about new responsibilities. This is probably the most important change. And responsibility not just means being there as a human, actually taking blame and taking responsibility again, if something bad happens.
And of course, to be able to do that efficiently, you have to predict all the ways things can go wrong. And well, that would require, again, a completely new set of skills.
So again, watch this space, there will be a lot of interesting developments and a lot of opportunities to grow professionally, I guess. Great final words. You said watch that space.
We too, we will watch that space and we will continue analyzing what's going on there. As usual, when we close on such an episode, we are living in challenging and fascinating times. We can really watch things change while we are working. So we will follow up on that. And as usual, for the audience that is watching, listening to this, if you have any questions, if you have comments, if you completely disagree with us, please leave your comments in the comments section on YouTube or send us a mail or leave it wherever you are and which platform you're using to listen and to watch this video.
Alexei and I, we are really keen on learning your perspective on this topic and any other that we're covering. And we will come back and have a new look, a fresh look at what's happened in the meantime. I promise you, no AI agents will be involved in that. We will be talking to us humans, guaranteed. At least we tell you so. So thank you very much, Alexei, for being my guest today. That was really a challenging and interesting conversation and it is not at an end. We will continue it and looking forward to that.
Thanks, Alexei. Thank you. Have a nice day. Bye-bye. Bye-bye.
