Cyber security used to be about keeping hackers out. Now it's about keeping your own AI honest. As artificial intelligence reshapes everything from coding pipelines to customer service, the attack surface has exploded and so has the complexity of who or what is responsible when something goes wrong. As CEO of Palo Alto Networks, Nikesh Arora has spent years building the tools that protect the world's largest organizations. But the game has changed. LLMs can be hijacked. Agents can be weaponized. Deep fakes can drain $20 million from a company in a single Zoom call. And boardrooms are now on the hook for the rogue actions of AI systems they barely understand. In this conversation with Suraka P at the India AI Impact Summit, Arora breaks down exactly what it takes to secure an AI-native enterprise, from firewall wrapping LLMs to building an agent registry for a workforce that's 80 to 1 non-human. He also makes a sharp case for why India needs a sovereign AI model and what the real 10x opportunity is for Indian cyber security startups. Listen on. It's Friday the 20th of February. I'm Aner Banchhri. This is the Morning Brief.
So as far as Palo Alto Networks protecting AI-native systems is concerned, what is the role that the company plays in protecting population-scale systems like UPI, Aadhaar, and what is the role that the company plays?
Well, from an AI perspective, it's a constantly changing field right now, right? Because LLMs create more and more capability every day. They started off as LLMs that allow you to ask questions and do creative things, and suddenly coding became very big. So now suddenly you have LLM coding. So you have to figure out the developers who are using the coding platform, how to protect their endpoints to make sure that rogue malware doesn't get on there, how to make sure that the code that is produced by the LLM is not malicious, to make sure that code is tested. So there are many, many aspects as AI gets deployed both from an LLM perspective and a coding platform perspective. And we have capabilities that will make sure we can safeguard the LLM, that nobody can hijack the LLM, nobody can take it over, nobody can promptly inject in the LLM to get you bad outcomes. And we have a capability where we can put a firewall around it to make sure that the LLM and all the AI artifacts are protected. And last but not the least, you also have a technology that constantly tests the LLM to make sure it hasn't changed behavior, because these things have behavior over time. As long as it's the expected behavior, it's fine, but you don't want unexpected behavior where it's being taken over by rogue agents.
You've said that 2026 will be the year when executives will be held responsible for the rogue actions of AI agents. So, how do companies and boards protect themselves from such lawsuits?
The biggest challenge is now that we're going towards an agentic future. There are two kinds of agents: there are agents that are delegates of you and me, in which case we're responsible for their actions, and that's what's going to happen. There will be a lot of delegated agents that are going to be created. They're going to be held at the accountability of the human level. The challenge becomes when there is no human responsible for the delegated agent, right? You've decided to give it agency. It can act on its own behalf. I think the best example is Waymo, right? In the US, you have a car that drives you around. There's no human involved. You go, it stops, it picks you up. Now, who's responsible if the car hits somebody? Who's responsible if the car has an accident? Right? Then that accountability falls on the enterprise, which means the enterprise has to ensure that that agent has been tested so many times that you're willing to give it agency. But I'm pretty sure that it's going to take a while before we have to train every agent in that environment multiple times to make sure that that agent is trustworthy.
So currently agents outnumber humans and the ratio is something like 82 to 1. So when the primary user is no longer a human, how do companies exactly rethink zero trust architecture?
Well, look, at the end of the day, every agent that works on behalf of an enterprise has to be registered somewhere. Just the way employees are registered in the HR system, there must be an agent registry in the future where every non-human agent of yours is also registered in that system. Every agent is going to have some credentials. Every agent has credentials and tasks. If you're a good enterprise, you have to make sure that each task that you have ascribed to an agent has been well tested and the agent is, you know, what the outcomes are going to be of that agent's task. You understand all the permutations and combinations. So unless you do that, you're going to create large liability for yourself.
So while global executives have deferred on whether India should be building foundational models, you've been clear that India should. So what is the role that sovereign AI plays in cyber security?
I think there's two questions. I think if you look at the evolution of foundation models, if you look at the amount of money that's being spent, it looks like it takes tens of billions of dollars to train the next generation of the model. If you look at, there's actually three large companies that are doing that sort of activity and they're spending tens of billions of dollars and raising tens of billions of dollars. So if you read the math, it says $650 billion are going to be spent in capex in the US from technology companies, half of that to deploy these models and majority of that to build and train these models. So if one wants to be in the models race, you have to make sure you're willing to commit that kind of capital, right? The idea that you can build models with a lesser amount of capital, maybe you can make 20% more efficient, 40% more, highly unlikely you're going to be 2,000% more efficient. So you have to still dedicate billions of dollars. That's one approach. The other approach is saying you can have task-specific models and there are certain models you don't need a genius to do certain set of repetitive tasks. So maybe you can have an older version of model or a smaller model that can solve the problem. I think more than likely many nations will apply a hybrid strategy where there'll be a sovereign model which can do a certain set of tasks which will be cost-effective and right for the market and train local domain knowledge, and there's more than likely you're going to need a large language model from perhaps one of these three companies if you're trying to do drug discovery.
So we have a wide developer talent base but there is a shortage of high-end AI talent. So how do we leverage AI as a service to bridge this talent gap?
Well, the talent gap is global, right? Even in the US, we have a large disparity between people who know how to build AI applications and use applications. So the US application you see is Darwin. Everybody learning themselves how to use Codex, how to use Claude Code, how to use Legal AI. So a lot of the younger people are applying themselves and learning that. The paucity of talent is in people building applications because there's no course that you can take in university which teaches how to do this. So literally all these founders, whether you're building Harvey.ai or Lego.ai or you pick your favorite AI application company, are self-learned. They've actually figured out how to do this themselves, how to build the capability. On the other side, you're seeing a bunch of AI researchers who've been training on AI for tens of years in the past who are helping build language. So you need the geniuses of AI at that end. You need really smart people who are entrepreneurial to build applications, and the entire populace has to be trained on how to use AI. I'm sure you learned how to use Frogs.
