Your AI Agents Will Find the Identity Shortcuts

Hey everyone. Welcome back here to Techstrong TV. My next guest, he's back with us, is Roi Katmor.

Roi is the co-founder and CEO of Orchid Security. Hey, Roi. Welcome back to Techstrong TV.

It's good to see you. Always happy to be here, Alan. So Roi, look, not everyone watching this has seen you on here before or is familiar with you or Orchid, so I don't want to take a lot of time because we have some good stuff to talk about.

But give people kind of a quick rundown on you, your kind of journey, and more importantly, Orchid Security and your mission and what you guys do. Sure. Happy to.

So, my name is Roi Katmor. I guess I've been in this space for too long. Not my first rodeo.

I came from the network space, moved to the endpoint space, and somewhere in the middle, did some fraud and others. Today, in the identity space with Orchid, who is basically an identity fabric for the AI era. We have to say that.

And in very short, Orchid operationalize identity, which means we discover, observe, and orchestrate basically any identity within an enterprise, and that means human, machine, agentic, anywhere it may reside, SaaS, premises, and cloud, put it into one place, provide you the insights and the understanding of what is the priority, and then actually can fix it with a click of a button so you can have a full scale and your foundation confidence for AI as well. Excellent. Just before we jump in, the website for Orchid Security?

security, as simple as that. Absolutely. All right.

So Roi, every day it seems we hear another report that some AI agent went rogue and tried to break in or tried to enter or whatever you want to call it, another third-party network organization. And it's a funny thing, right? Some people blame the agents.

Some people blame the companies behind the agents. I mean, at some level, blaming the agent's like blaming a dog. A dog does what it does because it's a dog.

An agent does what it does because it's an agent, and they have a mission, and they're going to try to fulfill their mission. I'm of the school that don't blame the agent. Don't blame, what's the term?

Don't blame the- The messenger ... the messenger, right. Blame the person who may have sent the message.

But when we talk about agent use in enterprises, right, these enterprises, as I said, don't blame the messenger. It's really the enterprise's responsibility to keep their agents in check. Easier said than done, but this is something Orchid's helping with, right?

It's 100%. I'll try to connect the dots together. There was a lot of, you just need to pick a week, but we picked a week after FortiBleed came again to the news where there was a big leak of the administrative credentials of Fortinet management and gateways.

And that basically illustrates the asymmetric in what we're trying to do, and I'll connect that to agent. Think about it. What's going on there, and why we're in this mess for two years, and how is it even connected to agents, right?

Think about it. For two years, they're trying to block something that already was exposed. And again, we see reused credentials again and again, and they urge customers to check and make sure that it's not existing.

And two years means that customers are struggling to actually contain because at the end of the day, and Orchid was in the middle of this, when we came to this and tried to solve it, we understood that the identity dark matter, those access that are embedded within application, that sits outside of the big central identity solutions and out of control. Basically, think about it as the local accounts, as the service identities, the API credentials, those path that bypass the enterprise identity, and you just don't know where they are. So the asymmetric is, an attacker can have a one credential.

You, as an enterprise, likely know only 43% of those usages. How do you make this asymmetric? Think about it.

An attacker needs a one stolen credential. The enterprise needs to find all places that this thing can work, and then you understand why is it so hard to contain. So how does this have to do with an agent?

" I'll give you some statistics. In 100% of our enterprise accounts, there were plain texts. Why do we leave plain texts?

It's like the basics. Because somebody wanted the app to be backed up at night. Makes sense.

So what do we do? We give it to some Amazon or Google or Azure or Microsoft Azure credentials and go and do it overnight in the maintenance window and throw it all there. So you got yourself now a credential.

It wasn't a concern before agents, because everybody said, "Don't worry about it. We contain it. It's segmented.

" But now we connected an agent to it. So there is something connected now, and it takes it, and we can talk more about that, nine seconds to elevate itself, which means what we're seeing, the hugging phase. All of those are controls that we're just not ready yet to securely adopt AI.

And so we left something behind. We gave it the front keys to do one thing, but we got an ambitious agent. It want to please us.

And it found a plain text inside, and now it elevated itself in nine seconds, and we're completely blind to it. So dark matter, containment issue, and agents are what we're seeing right now. And yes, you can blame the messenger.

We need to be ready and not just adopt. And I want to emphasize that. Nine seconds, that's an eternity for an agent, right?

" But again, the agent, let's not infer evil intent to the agent, right? The agent thinks it's doing its job, right? It's going to do it better, faster.

It's going to fulfill your wishes. So, the flip side is, how do we with a forethought lock this down so that the agent stays kind of in the box that we put it in? It's a very good question.

And as you can see, it is complex. But I think, I'll give you a good analogy, and then we'll walk through the blueprint. I always put it as an analogy as a painting.

I don't know if you, the ones that are maybe more gray hairs, still remember how it is to actually paint, and I always say it takes 30% of the time to paint and 70% to prepare for the painting. And if you haven't prepared well, well, you know how hard it is to get a little stain from a wooden floor, and that's exactly the same with agents. The preparations, it's so easy now to take a studio, right?

Like you take a brush, and you write yourself something, and everybody's super proud of what they're doing, but they also don't understand we didn't prepare to it, and now we got a lot of stains to remove. Let's talk what are the stains and how does this come together. We need to understand the pieces that we don't know.

And the pieces that we don't know are the identity dark matter. Those local accounts, everything that is outside of the identity and access management stack is something that needs to be aware of. And I'm not saying go fix it.

I'm very realistic. But be aware of. The observability of your identity is a key to safely adopt AI.

Once you know and have an audit and understand the picture, and I'm sure the picture is not going to be the greatest. We didn't paint yet, okay? So it's not great.

At least you know what can be connected to AI and what cannot be connected to AI. If you're ready, ready means your hygiene is well established. You have an audit.

You need to know what effectively, not by design, there is an intention and there is an effective. Make sure that they are aligned. Audit will be your best friend in knowing.

You need to have your controls in place. Everybody got the right solution in identity. The problem is not the solutions.

The problem is onboarding into the solution. A sale point is as good as the data you provided. And that's the reason.

So we said, be aware of your locally governed, so your identity dark matter, have an audit, have your controls in place, and know where you stand, and then you can rank it and say, "Well, here I am AI ready. I got an audit, I got full control, and I got full visibility of all the different authentication flows. " Perfect.

You're ready. Now you can put it out there and test it. And that's going to be, again, going back to the more of a whitelisting rather than let it run because removing the things after are going to be very hard, especially in a place with a dark matter and you really don't know where it is anymore.

And that is something we're seeing already. Got it. All right.

Let us turn now to Orchid specifically. I think we've defined the problem. You talked about a lot of ways that you can help solve it.

What specifically does Orchid do to help us here? So let's go back to our little starting point. Orchid first discovers all actors anywhere they reside.

That means human, non-human, and agents. There's a reason why we kind of broke it into this. " Don't forget, agents eventually work.

They are triggered by humans, created by humans, using NHI credentials, and there is a full story around this. Orchid can bring all these three dimensions together, all the actors into one place, but can do this for anywhere they may reside. And that means you, as Alan, maybe ran your own agents on your own computer to do some things for you.

Maybe the enterprise gave you an agent to your service. Maybe your SaaS vendor gave you some agentic to run on theirs, and all of them actually live different lives. One of them is on premises, one is on the cloud, one is in the SaaS.

Orchid can actually bring them all together to one place so you can operationalize it, meaning you're going to go now, in a day-to-day, we talk about agents. '" And we're going to say, "Wait a second. " It has a plain text for the sake of the argument.

You connecting this, you still have an agent, but you have a different permission that what effective permission than what you intentionally thought you're going to have. Orchid can flag it, Orchid can tag it, and Orchid can actually put kill switches to say, "Okay, Alan, it's drifting. You did it despite the fact it wasn't ready.

" So for example, it could be block it, totally fine. Kill the agent. " I love that.

Yes, if this is what you want, maybe that's the tool that you need to use. " Maybe it is. But eventually, human in the loop or stopping that specific or forcing the specific workflow or taking it down altogether are valid options.

Just make sure you pick one that is valid and to the adherence of your compliance and security practices. Excellent. Roy, we don't have much time, but I did want to bring something else up.

I was reading an article this morning. We seem to have, in AI time, every two weeks is a new century, right? Wow.

So just when we started wrapping our heads around using Claw and OpenAI, OpenClaw, and all the permutations of that, now we see a new class of agents like Muse, Dot, and any other. They come up with one-word names now. They have cute names.

Does this represent a different class of agent, a different way you got to deal with it? Because they're more persistent now, right? They hang around, and they keep at it.

What does that mean for agent security? What does it mean for Orchid? Look, I think we see two trends.

On the one hand, we started with nobody knew two years ago how enterprises are or will adopt agents. I think that there were many different ways. People thought that all of us are going to have studios and we're all going to build what we need.

And we came out to the reality check that we, like in DevOps, like in data science, we're going to have an AI team that's going to get orders and going to do this for us so we can actually have compliance and other security controls in place. That's a good thing, but it's still disconnected. Now we have another team that needs to work with other teams, and that's not exactly, maybe not the request that you wanted, and that's not maybe not what the app has to have, or the app owner has a different intention, or the app is not ready for it, and there is some disconnects.

We've seen this in the beginning, and we saw a lot of hugging faces, and we saw a lot of those type of jailbreak. And again, misconfiguration, dark matter, readiness was always the case. What we're seeing now is actually a different level of agents.

The agents are becoming more and more autonomous. The trial and error is something that is happening, and the agent is becoming more and more advanced in actually trying to pursue what it needs to accomplish. And on the one hand, that's great.

On the other hand, nobody's, again, checking the guardrails and make sure that we're ready. But you can see the other trend of it, for example, from the infrastructure teams, I don't know if you had a chance, NVIDIA, OpenShell, putting very hardening on the boxes to make sure they're not breaking out. It's true that it's just on the box, so it's not that the agent can't do other things in other places.

But to some degree, I think the trend is we see more autonomous, we see more advanced and improvisation in an attempt to accomplish a deal, I mean, a task. But on the other hand, I think we see the vendors, the frontier models, and the chip manufacturing behind this starting to understand what's going to happen and how that would slow them down. And most likely speculative, of course, is pushing them to putting more and more controls, understanding that we'll see more and more autonomous.

And now it's a race between are we going to be able to control it on time or the next autonomous is going to get out of control? Because, again, not killing the messenger, but we weren't ready, and we'll see another big headline. And hopefully, the headlines are not going to be too big before the controls are going to be in place.

Good. Roy, we're over time. I got to end it.

But thank you so much for coming on, giving us a little education. You have a wonderful way of describing it. I think people like that.

Continued success with Orchid. Come back and see us soon, okay? Absolutely.

And thank you so much. Great seeing you again. My pleasure.

Roy Kaplur, co-founder, CEO, Orchid Security, here on Techstrong TV. We're going to take a break. We'll be back.