00:18:52 · Nikesh Arora, Aditya Agarwal
I've done that here for eight years now. [00:18:55] S1: It's been eight years. [00:18:56] S0: Mhmm. Almost. [00:18:58] S1: So that's been interesting. So eight years, obviously, we had the ChatGPT moment coming up on four years now. So what was the overall AIobviously, you guys had a ton of machine learning at Palo Alto Networks. It was a key part of your overall technology stack. What was your AI posture prior to generative AI coming into the larger consciousness? Was that something you were working on in a more research setting? [00:19:25] S0: Mostly, it was machine learning. [00:19:27] S1: Think Can you talk a little bit about So your stack
00:19:31 · Nikesh Arora
[00:19:31] S0: what's fascinating is I still believe that many labs, whether it's Google or perhaps I don't know if Meta or not, but other labs many labs had the capabilities that ChatGPT had when it came out. I think most of them didn't believe this was a real product because if it was hallucinating 20% to 30% of the time, how could you make something useful come out of it? I think they anticipate or understand the consumer use case where we as consumers would make our own judgment and we could tell how much of that was hallucination. In some cases, we didn't care because writing a Shakespeare sonnet and Eminem style is cool. It doesn't matter if it's right or wrong. And they didn't quite understand the value of that product. And I think that's what Sam Altman did with ChatGPT and came out. And that's kind of flipped the bit. And that is why we're all here trying to start companies as quickly as we can. I think that's kind of what gave the impetus to generative AI, at least. Until then, because of the hallucinatory nature or the nondeterministic outcomes, none of us in real business applications were comfortable deploying that technology where we couldn't predict what it was gonna do. I think subsequently, everyone is gingerly, at least I'll call it the incumbents, which are us versus the new companies you guys are gonna build. All of us are gingerly infusing LLMs into certain use cases where we believe they'll do a better job than what we were using in the past, perhaps not as much as we need to. I think all of us are slowly rethinking workflows in businesses to see how business is gonna be conducted in the future. And in the short term, it's slow going. I do believe longer term, it's gonna get better and probably more than any of us can imagine.
00:21:20 · Aditya Agarwal, Nikesh Arora
[00:21:20] S1: Perhaps I'm giving you a layup here, but obviously, we'll talk about Mythos in a second, which I think that the cybersecurity positioning as it relates to AI has dramatically changed over the last even just three or four weeks. But prior to that, how much of a tailwind was kind of essentially generative AI to your business or even to your technology stack? Was it fundamentally changing both what you were doing, how you were doing it, and in some ways, how you were able to charge for it? [00:21:47] S0: What's fascinating is in our industry, in cybersecurity, we have to think one step ahead of the technology adoption curve. Because when ChatGPT comes around or LLM's come around, we're trying to figure out, well, how are people going to use models? How many models on average are they going to use? How are you going to inspect the traffic of the models? There's prompt injection. There's data poisoning. There's whole kinds of bad techniques that are out there, or bad actor techniques that are out there, that will be used to take over, hijack these models. Given their learning mode, their nondeterministic mode, their bidirectional traffic mode, a whole bunch of new cyber capabilities had to be built to deploy them well. So we were constantly trying to figure out how this thing's going to work. And then you throw agents on top of that. Then you throw these MCP servers on top of that. None of these things have security built in. You told me about open claw trying to [00:22:36] S1: My deploy bad. Yeah. It's a nightmare. Terrifying. Yeah.
00:22:36 · Aditya Agarwal, Nikesh Arora
[00:22:36] S1: My deploy bad. Yeah. It's a nightmare. Terrifying. Yeah. [00:22:38] S0: So I made the joke. TSA was not invented when they made planes, just FYI. You guys might think so. When they hit planes, they were just trying get from one place to the other. There was no TSA. TSA came about much more recently in our lives. So security is very often an afterthought because you suddenly discover the bad actor can use them for bad things. So in our job, we are supposed to make sure that we anticipate the bad actors' behaviors, build technology in anticipation of how you're going to deploy it. Most recently, a large company got hacked. Many large companies got hacked because one of their agents got compromised. Now they want to put security for agents all over the place. I'll tell you what. How the people are still struggling on the definition of agent, let alone agentic AI security? What is an agent? We can have that debate afterwards in coffee. So there's so much stuff that's going to happen. So we're constantly busy trying to figure out how to build the scaffolding, how to build the control points, how to build the kill switches for when AI goes rogue, which it will, it has, and will continue to do so for long. So we're busy fixing that stuff. I mean,
00:23:42 · Aditya Agarwal
[00:23:42] S1: it's crazy, right, even before you look at... And we'll, again, we'll get to meet those in a second. To your point, all of these new parts of the stack that have been introduced over the last three or four years, they have zero security in them. And arguably, we haven't even seen like... Perhaps you have seen it, but we haven't seen a major attack through one of those kind of vectors, whether it be an unsecured MCP server, kind of a weird prompt injection attack through a model. We are hardening ourselves against all of that. And then we'll obviously get to these new coding models getting to the point where their Mythos is exposing a huge class of vulnerabilities. So does that mean I need to go out and buy like some Palo Alto Networks stock?
00:24:23 · Nikesh Arora
[00:24:23] S0: Well, we could do whatever we want. It's for sale. That's what I'm told. But before we get there, look, every new technology, when it comes about without security by design, there's tons and tons of different sort of places that today, there is no secure version of MCPs. And I read yesterday or two days ago that Salesforce went headless. Well, I don't know what the security governance controls are because not too long ago, one of their marketplace apps had gotten hacked and a whole bunch of companies had to go deal with that. Now just imagine, you're going to be running agents talking to Salesforce agents, all kinds of stuff going on. There's no authentication. No company knows how many models they have in place right now. On average, if you go to Hugging Face, the people are downloading a million models a day or a million downloads a day. Going to some enterprise environment. Nobody has any idea how safe those models are, how secure they are, if they have backdoors or not, whether there's a backdoor connection that is sleeping that opens up three months later. You don't know. You have no idea what's in the model behind it. So there's tons and tons of security vulnerabilities as it's related to AI implementations. There will be tons and tons of debate, discussion, and I don't think there's a standard architecture to deploy agents. If I asked you, you work for an enterprise, or how do you create an inventory of all agents that you have? The answer is I don't know. That's a great place for us, by the way. If anytime customers say, I don't know, that's when we show them, this is how it's gonna work.