The Agentic AI Governance Gap: The Security Reckoning Ahead

We're live. Okay. Hey, welcome back here to our...

It's actually day two. Some people may call it day one. It's Tuesday.

We're at RSAC Conference, continuing our coverage. My next guest is Greg Clark of OpenText. " Well, I forgot to mention, Greg is with OpenText, but he's a, was it- Senior director ...

senior director. I was going to say- Product management ... I almost gave you a raise there.

Well, hey, field promotions are- That all I said ... almost as good as the real thing. Exactly.

Senior Director of Product Management for cybersecurity at OpenText. Greg, we were talking off camera, and you said you've been around the carousel with OpenText now two or three times. Yeah, exactly.

Give people a little sense of your journey. So, again, I started my career at OpenText maybe 20-some odd years ago. It's an information management company.

So I started in content management. About halfway through my career, I pivoted into regulatory compliance, eDiscovery, which led me into data security. And as we started to evolve our data security business and put together a portfolio that was cybersecurity centric at Micro Focus, we pulled together identity and access management, data security, SecOps, the intricate pieces that make a cybersecurity portfolio.

And then three years ago, we were acquired by OpenText, and I'm back for a- I'm back ... I'm back for a third time. Just when you thought you were out, they pulled you back.

Exactly. Good for you, man. Exactly.

Good for you. OpenText is such an umbrella of an organization. " Right.

But what do they do? Software, this, that, the other thing, but do they really know what they do? Well, they may know some of the things.

Correct. Give the audience, if you don't mind, an idea of what do they really mean. So, as I mentioned, Alan, we talked about information management.

In today's environment, especially if we're doing the buzzword bingo around this place, around agentic and around AI- It's true ... the concept of secure information management is what our customers really entrust us on a daily basis to do. So what do I mean by secure information management?

What we have is GenAI, agentic AI. It's outpacing the governance models that our customers have. So, the bull is out of the barn, so to speak.

So when we look at our capabilities to help organizations secure that information, to secure the identities that are accessing that information, but not just human and non-human, but agents. They need to be on the same playing field as others because they're 20 to one, 50 to one when they're deployed inside of customer environments. But you also need to monitor that behavior.

So as agents are traversing data, as users are traversing data, look for unusual behavior, look for potential risks. So when we bring all those pieces together, along with our information management background, we have a secure information management platform or capabilities that are built exactly for enterprise AI, or what I would consider safe AI. So the ability to manage the data, govern the identities, continually monitor behavior around the information.

And also, on the application security side, we're ensuring that the code is secure, that agents, as they're onboarded, built, or bought, are legitimate, verified as well. Got it. Greg, it's interesting that securing an API, let's say.

Mm-hmm. Look, the API was written to do this, and it does this every single time. Yep.

So you can make rules around that- Yep ... because you know this is what it does. Yep.

It's deterministic. Exactly. When we're talking about trying to create rules around agentic AI, the game changes a little bit.

It's non-deterministic. The agent learns. Its behavior patterns change.

As a matter of fact, I've been playing with agents now for a month or so. It doesn't do the same thing twice. I'm constantly like, "Hey, this is different than yesterday.

" The model changes. And it says, "Oh, you're right, Alan. " Yep.

Exactly. But it's hard to build models around that. It's hard to build the guardrails around that when this thing is zigging when you think it's zagging.

Yeah. And so how do you account for that kind of, I don't want to say randomness, but it's kind of random. Yeah.

I think there's a couple foundational pieces, really, of note. So, if you look at one of the key concerns, so what's the attack surface? It's the data.

The bad guys or insiders are trying to get at data. So, at OpenText, we have the ability to help understand and visualize risk around data, and that's one aspect. SoWhen we are able to visualize that risk or surface that risk up into the identity layer as it sees unusual behavior into the SOC, as it sees unusual behavior, we're able to heighten the signal, raise above the noise that, oh, okay, there is consumer data in this CRM system.

There is a SharePoint site that has sensitive data. How do we help see through the noise? So not only can we show the risk visibility, but when you talk about agentic and how it's scaling beyond the confines of what we would consider, an API talking to an API, a person traversing the environment, our ability to protect the data at the source with encryption and format-preserving encryption enables us to essentially future-proof data- Okay ...

so that before it goes into the AI pipeline, it's de-identified. So Alan's sensitive data is scrubbed, but just like an analyst would use it in a cloud data warehouse or a BI tool, as an agent is accessing that data, it still has referential integrity. It can still do its task, right?

So the guardrails around data is one aspect. The other aspect is when we start looking at agents as an identity on the same plane as a service account, as a non-human and a human, we start putting up the abilities or capabilities around what is the intent of this agent? What can it do?

What data should it have access to, and what are essentially the policies for this agent? The moment we start to see drift, where an agent is talking to another agent, it's accessing data, not in this CRM system, but a layer down or in another application. You detect that drift with this continual monitoring, you're able to, again, step up authentication even for the agent, just like you would a user.

You could block the account. Sure. You can suspend the- No, at that level, you can.

Yeah. It's just another user. Exactly.

So those are foundational pieces that we can bring to the market for customers, so that they can adopt AI more reasonably, responsibly, and from an enterprise AI standpoint, those are the things that are absolutely critical- Mm-hmm ... for these things to scale. Right?

Love it. Yep. Love it.

Let's pivot a little bit. You guys announced, I think it was just yesterday- Mm-hmm ... a study you did with the Ponemon Institute, Dr.

Larry Ponemon. Nice. Been dealing with his stuff for a long time.

Yep. Talk to us about that, Greg. Yeah.

So it falls in line with a lot of the things that I just described, Alan, but a couple things I think that are what we've seen. So this was 1,900 respondents at the VP, senior level with AI as a clear mandate for them in terms of their priorities. What we found was in that report that, look, about half of them had deployed gen AI, but really, 20% of them saw any material success or value out of that.

So it really laid out the foundation for what's next. So when we start looking at autonomous behavior, even less of those respondents, 43%, were confident that AI was ready to be autonomous, ready to make decisions, ready to take on tasks. But the biggest thing I saw in terms of what I felt was surprising is about 60% of the respondents felt that regulatory pressures, privacy, compliance were impeding their ability to really gain some scale and some momentum, live up to the promises- Absolutely ...

that agentic AI has. And again, it's one of those things where what I described around governed identities, protected data, continuous monitoring, really understanding agents as they're built and bought and deployed, those are the kind of guardrails, or those are the kind of capabilities that we believe are the main DNA behind what is successful and safe AI, enterprise AI. Absolutely.

Let me ask you a question. Anything in the report that you sort of didn't see coming? Or maybe, I think what we see with AI when we look at these reports is, I saw it was coming, but I didn't think it was that far along, right?

Right. I thought a lot of people were using it- Mm ... but not 88%, right?

Right. Anything like that? Yeah.

I think the thing that stood out was we had, 18 months ago, we were talking about this inflection point, and we felt that, look, people would learn from the gen AI journey, and they wouldn't let the floodgates open up. So what stood out to me was actually how many organizations had, yes, deployed agentic AI, but also don't have enough risk-based approaches to monitor and guide the successful deployment- Got it ... of these things.

So we didn't learn from gen AI. I think I said on stage a couple, maybe a year ago, while we certainly won't be making the same mistakes with agentic AI. I've been around long enough to know better than that.

Yeah. I've been around long enough. And so have you, to tell you.

I should. Right. Fooled you.

Yeah. Exactly. Got you that time.

Exactly. But the fact is, look, not saying AI is like things that have come before, because it is very different. Yep.

Some of the patterns- Mm-hmm ... of adoption, and especially some of the patterns of securities response to it- Right ... are kind of the textbook things we've seen over and over- Right ...

again. Right. And it's like, I don't know what it'll take for us to learn those lessons.

Mm-hmm. But I just hope we learn them a little faster each time. Well, the scale of AI is probably- The scale and the time, the velocity of AI ...

yeah, is the scarier part, right? Yeah. It won't be a chatbot that makes up a bereavement policy like in Air Canada, or you can trick a chatbot into selling you a Chevy Tahoe for a dollar.

It will be you're exposing your entire customer base, your CRM system, to an agent who has then shared it, or acted- Well, what I worry about is, are the agents hackable? Yes. Absolutely.

Every agent, every interaction, if you're not monitoring every agent, every interaction, every action, you put yourself and your data... And let's be honest, it's a reputation risk. Risk.

Yeah. Right? All on the line.

You're putting that all on the line. And then the other thing is these agents, they spawn sub-agents. You don't know what those sub-agents...

It's a complicated thing. Where can people get this report? They can get it from our website.

com ... on the front page, maybe And I believe it's right off the front page. But also, if you look at us in our socials, on LinkedIn, on X, the report will be front and center there for people to download.

I love it. Greg, anything else that we haven't covered? I think, Alan, you've got a great reputation.

This is where I'm pumping your tires. Yeah. You want to pump my tires a little bit?

In cybersecurity, this is a checkbox. Yeah, no, this is good. I'm glad.

Right? Don't be silly. I'm glad to have you here.

But let me ask, so it's only day two. I don't know how much you were here yesterday. We were busy.

Yeah. But what do you think... Certainly the story this year is agentic AI.

Yeah. Duh. Yeah.

But what else are you seeing under that surface? I think it's really around the importance of AI governance. Yeah.

I think that when you start to scratch a bit of the veneer off of it, the ability to discover agents, shadow AI, again, monitor them as they're coming on board, what their behavior is. Because look, there could be thousands of them. They could spin up for one workflow and then spin back down.

The AI governance side of this challenge that we're facing here is the next horizon. I think the whole governance of agents, of the code they spawn- Yep ... is really going to be the battleground, at least through the next year to two.

Badly. Until we figure how to... We need an equilibrium to develop.

Yeah. Right? Exactly.

Hey, Greg, thanks for coming on here, man. Oh, my pleasure. Good luck.

My pleasure. Look, you're out in the front lines on this one, man. Ready to go.

Greg Clark, OpenText here. Go check out that Ponemon Institute report on their website. Dr.

Larry Ponemon. We're going to take a break. We'll be back with more here at Techstrong.

We're live from RSAC at Moscone West.