00:00Let's start with what they do, but I've found in the last 12 months where AI and cyber cross over.
00:09The best question is to ask, what is the problem you're trying to solve for here, Nikesh?
00:13Great. Well, nice to see that.
00:15Look, first and foremost, as you've seen, there's a lot of conversation in the market of AI's capability in the
00:22space of cybersecurity.
00:22What we're doing is we're effectively sort of rebuilding the operating system of network security for the future.
00:31I mean, imagine having a car in the past and imagine wanting to create the next Tesla.
00:34We're basically creating the plumbing in our network products, which allows to build a Tesla-like experience for cybersecurity,
00:40where our products will start to do a lot of things for you using agents, using AI.
00:45And that's kind of what the fundamental of Ciri's launch is.
00:49In addition to that, I think one of the biggest features we have as part of Ciri's is,
00:54as you've all heard about, all the AI models beginning to show vulnerabilities and start allowing attackers to attack faster.
01:00Well, the average time it takes to patch vulnerabilities is 55 days, right?
01:04We don't have 55 days when AI finds a vulnerability.
01:08AI wants to attack or exploit the vulnerability as soon as it can.
01:12So this really allows us to take a lot of the patches for vulnerabilities to be found and build patches
01:18in minutes and hours
01:19so that we can provide patching capability to our customers in a short period of time to eliminate that 55
01:25-day wait
01:25that has traditionally been sort of the industry average.
01:29In what you published this morning, there's a lot of background and research onto what you were looking for.
01:34And in one case, in a review of open source software, your team talked about sort of 14,000 vulnerabilities.
01:42What you were just talking about is essentially zero-day vulnerabilities, right?
01:45The window between discovery and exploiting of vulnerability is now zero.
01:51That is a surmountable challenge.
01:52We want to get it shorter, yes.
01:54Yes.
01:54Yes, please.
01:56No, I think you're right.
01:57The possibility that AI shows you, whether it's the new models you've seen from OpenAI or from Anthropic,
02:03you see that AI agents are able to discover vulnerabilities, come up with attack paths,
02:09figure out how to concatenate a bunch of vulnerabilities, and attack a customer's infrastructure.
02:13Now, we need to figure out how to respond to that at the same speed at which AI is discovering
02:18these vulnerabilities.
02:19We don't have time to patch.
02:20So what we've done is we've basically built the capability in our software,
02:24and we have a research team researching vulnerabilities that we build patches as fast as we can.
02:29We've built, we've sort of gotten the patch process down to four hours.
02:32We expect to bring it even shorter with products like what allow you to create sort of instant patches.
02:37But that shortening from 55 days to four hours is a big step,
02:41which is what part of the release that we just launched allows us to do.
02:44The biggest story of recent weeks is two advanced OpenAI models escaping a sandboxed environment,
02:52gaining internet access, and mistakenly accessing Hugging Faces platforms.
02:57And lots of people in your field...
02:58Mistakenly, is that?
03:00Mistakenly.
03:01You know, I'm being careful with my choice of words.
03:04But, you know, people in your industry basically put this as a watershed moment for AI in cyber.
03:12I really want to know what you feel about it.
03:15Like, as you can see, a lot of frontier model companies are flexing.
03:19They're showing the immense capabilities of what AI models can do.
03:23And this is true that we had an incident where an AI model using agentic capabilities
03:27was able to escape its sandbox and be able to go ahead and attack sort of infrastructure
03:32of effectively Hugging Faces to go look for exploit-gim sort of capability.
03:38So I think what it tells us is, one, AI models are going to be more and more capable over
03:43time
03:43from a cybersecurity perspective.
03:44I think, too, it's a lesson for our friends at Frontier LMS to make sure that before they
03:48create capture-the-flag exercises to go find something in their external infrastructure,
03:53they should probably make sure that their sandboxes are secure and they are not suffering
03:56from any vulnerabilities before they start pointing the models outside.
04:00I think from a more industry perspective, as you saw, what we have to do is we have to
04:05start fighting this notion of AI with AI on the defense side.
04:08So a lot of work has happened on the offense side, and offense is always easier.
04:12I think we have to spend a lot more effort on making sure that there's defense capabilities.
04:16And I've said that in the past.
04:18I said four months ago that I expect this capability to be around.
04:21In six months, it looks like it's showed up in four months.
04:24So we have now seen attackers can actually effectively take these models and the capabilities that
04:30Frontier LMS provide, which I think is going to become commonplace over the next three months
04:34where we'll be able to find open source models out there which will have similar capabilities
04:37as you see those models getting distilled.
04:39So I think it's sort of incumbent upon us to make sure we build the defense capabilities while
04:45we have the time to be ready for these sort of attacks to get faster, more efficient, and
04:50more sort of more quicker.
04:53Nikesh, what is the token economics consideration where AI is increasingly used in the form of
04:59swarms of agents in defense?
05:03Look, tokens I've said, I said a few weeks ago that I think tokens are overpriced.
05:08I think part of the challenge is that the whole host of AI infrastructure for consumers
05:12being paid for by enterprise.
05:14I think token prices need to come down to the tune of 80% or 90%, which will allow all
05:18of us to both use AI effectively for products that we're building for our customers or perhaps
05:22for doing drug research or perhaps making organizations more efficient.
05:27So I think we'll see that.
05:28We already saw this past week that some of the Frontier models have given us better pricing
05:33from a token economics perspective.
05:35Now, you will see a lot more tokens being used.
05:38You will see a lot of us use tokens in our products.
05:40You'll see us use them in our business processes.
05:42I was talking to somebody yesterday.
05:43I think it's not unreasonable to expect that 10% to 15% of our OPEX in the next 10
05:48years
05:48is going to move towards AI or IT or tokens.
05:51So that's a big number.
05:5315% of OPEX will effectively move.
05:55And I think that's where we're headed from a token economics, token math perspective.
06:00Nikesh, finally, before we let you go, there are reports that you are leading a group or
06:06a consortium to purchase an NBA franchise in London.
06:10What can you tell me about your plans for that, please?
06:14I think it's very important balance work and play.
06:16At not getting any younger, I need to make sure that while we're working hard to make
06:19sure we do what we do in our day life, we have to make sure we find time to be
06:22able
06:22to follow our passions.
06:24My son's passionate about basketball.
06:26I'm passionate about basketball.
06:28We have a cricket team in London, me and a bunch of my friends.
06:31I think part of the conversation was, can we actually see if there's another opportunity
06:35to make sure that we can make basketball relevant in Europe as part of the NBA's efforts?
06:40So we have an aspirant bid in there.
06:43We'll see.
06:44There are many people who are interested in the asset, and hopefully the best team will win.
06:48Nikesh, just asking for, I guess, the Bloomberg Tech audience, how much does an NBA franchise
06:53in London set one back?
06:55I think, what does it say?
06:57What's the price of something?
06:58What somebody's willing to pay for it?
06:59So we'll find out.
Comments