Out of all corporate uses of AI, you’ve said it seems to be lagging in cybersecurity. Why?

Patel: There’s two reasons for it. One of them is the efficacy with AI tends to be low and the costs tend to be prohibitively high. And then the skills shortage has been pretty profound so far. People are just trying to just keep their heads above water. 

On the efficacy side, what do we think needs to happen? If you look today, most of the AI work has been done with generic models that have also been used for a multitude of other areas. I feel like there is an opportunity for having a much more specialized bespoke model or set of models that create this notion of what we call super intelligent security. We launched this at [springtime cybersecurity conference] RSA, where we said, ‘Let’s make sure that we have a security-specific model that eventually can do reasoning, that we should open source, that ideally is going to have a very different set of performance characteristics, from both an efficacy perspective and a cost perspective.’ 

This model was trained on an 8 billion parameter model, which is small. It’s showing in early benchmarks that it’s performing [as well or] better than a 70 billion parameter model. It can be run on a single A100 GPU versus, having a cluster of 32 H100s, which is what typically a 70 billion parameter model would need. Why is this important? A huge constraint is having the infrastructure to go out and use AI effectively. The cost is typically pretty high if you go to a company that’s not in the tech space and is a low-margin business. It’s very hard for them to go out and start putting millions of dollars towards GPU costs. 

The second thing is we have to secure AI itself. The biggest risk we have is not AI taking our jobs away. The attack surface increases so dramatically that it could have harmful effects that we haven’t fully considered. That requires that you start from the foundation of saying, ‘I’m going to secure AI itself, rather than just use AI for my cybersecurity defenses.’ 

So what does it take to secure AI? You need to make sure that you have full visibility on what data is flowing into what models. You can’t really protect something you can’t see. No. 2, you have to then say, ‘Can we validate these models?’ Because these models, by definition, tend to be unpredictable. We are trying to build very predictable enterprise applications on top of a foundation that's inherently nondeterministic. We have to make sure that we validate to say, ‘Are these models working the way that you want them to work? And if not, then what kind of guardrails do you need to put in place so that the model doesn’t start behaving in a way that can then put you or the company in harm’s way?’ 

How much are businesses thinking about both of these things? I would imagine that the whole business has more of a purview on the first one, but are most even considering the second?

I think the use of AI has been increasing, but not quite at the base of the growth of AI itself. This is where I think there’s an opportunity. We do readiness indexes and CEO surveys to find out the sentiment of the market. There was this one stat that’s really stuck with me: 97% of the CEOs that we talked to were really excited about AI. Almost giddy excited, like, ‘Wow, this is going to change our business.’ Only 1.7% of them felt like they were prepared. There’s this big disconnect between people getting excited about the possibility, but not really having a level of preparedness. 

Then you double click and say, ‘Why is that the case?’ Three things come to mind. They feel like they don’t have the infrastructure knowhow to be able to fully prosecute the opportunity of AI. Second, they feel like the safety and security dimension, which is what we call the trust dimension of AI, is still low and people are worried about AI to really lean all the way in. We’ve got to solve that. The third one is they just don’t have the right level of skills, especially in cybersecurity. That’s a pretty rampant skill shortage right now. Three to 4 million jobs go unfilled every year in cybersecurity, and that problem’s only exacerbating; it’s not getting better. 

I feel like companies are thinking about it, but I think this is still far too complicated. Companies like Cisco are trying to work at simplifying the use of AI in the fabric of what we do.

What advice would you give to a CIO or a CISO who is grappling with this right now?

There’s only going to be two classes of companies in the world. There’s companies who are going to get very dextrous with the use of AI, and then there are going to be companies that really struggle for relevance—and you don't want to be the second kind of company. You have to lean in with AI, and lean in with making sure that people understand that the only way that adoption and AI scale is if you solve the trust problem. 

Unlike the past, where security used to be an impediment to adoption, today security might become one of the biggest accelerants to adoption. Because if you don’t trust an AI system, you’re not going to use it. But if you trust it, you’re going to use it. And in order to trust it, you need to make sure that there’s a common substrate of security that’s embedded across all of the models, all the agents, all the applications. 

Imagine when you start using these systems where you just ask an AI agent to do the work for you: I’m going out with my spouse for my second anniversary and I need you to book me a movie as well as a restaurant. Make sure that the restaurant’s not too far from the movie, and here’s my credit card. Go ahead and make the reservations and buy the tickets and pick the right seats so that two of us can watch the movie properly. In order for you to do that, there’s a technology element, but then there’s a trust element. Do I actually trust the system to give it my credit card number and autonomously work with other agents, exchange data and come back to you with something?

One of the reasons why this is so important is the majority of the breaches and hacks are because of identity. People are saying, ‘Why do I need to hack into a system if I can just steal the identity through social engineering and log in?’ That problem gets even more exacerbated as you have many more agents that are talking to each other. I have an orchestrator agent that’s talking to maybe five other agents. Agent one is booking a restaurant, agent two is booking a movie, and agent three is booking an Uber, and all three of them have to work well together and exchange data. Every agent is going to need to have an identity, and every identity needs to have a set of permissions that then knows what data to exchange with some other agent. And so identity is no longer going to be just about a human. It is also going to be about a machine and a service identity and an agent identity. 

In my mind, in the next year to two years, the progress of AI will largely be dependent on the progress of security and people psychologically feeling a level of trust with AI. That trust has to be real, where you have to have the right underlying technology to keep humans safe and secure. It’s not just about keeping us secure. It’s about making sure that we can unlock human potential if we keep ourselves secure. Now, what [security professionals] are saying is they’ve been made responsible not just to manage risk, but the degree to which you can unlock human potential will be based on the progress that the security industry makes on managing risk. Literally, adoption will accelerate, based as a function of security and safety being well-formed.