Name: CEO Initiative 2023: Unlocking AI’s Potential…Securely
Uploaded: 2023-12-18
Duration: 17 min 46 s
Description: CEO Initiative 2023: Unlocking AI’s Potential…Securely

CEO Initiative 2023: Unlocking AI’s Potential…Securely

Uncategorized

I would love to start with all of your perspectives on briefly sharing your view on the current state of generative A I and also the biggest risk that you think business leaders here today should know about. That's all. Well, I'll start, I think generally I, we heard this morning it's a general purpose technology that will impact every industry. I think the industries that will benefit and the companies that will benefit or be impacted. Uh Most in the short term are industries or companies that are digitally mature that have rich uh data sets and can move relatively quickly to take advantage of the new technologies in that spectrum. I think the industry that will accelerate probably the fastest unfortunately, is the cybercrime industry. Uh because cybercrime by definition is, is digital and it's largely actors who don't have to worry about, you know, ethical guidelines and safeguards and standards. And so they are able to move very quickly and we're seeing that already, we're seeing pretty much every company here is already being uh attacked using A I uh for fishing uh for, for malware, uh for social engineering with bots to try and find A way through, through your data and your application. So I think that will accelerate significantly with generating the A I with the ability to emulate humans and create much more intelligent social engineering and fishing attacks. So better than those emails I get about Alan needing gift cards, they're, they're going to get way better and way more intelligent. No, I think that what francois um outlined is, is, is probably the thing that in the short term is where the risk is. Um, but I'm actually pretty optimistic that we'll get past that. I think that the idea that we're verifying people by what we think are human traits. It's probably not a very good security implication in the first place like your banker shouldn't wire money just because you call up and you sound like you. And so I actually think that, yeah, that's gonna be a problem. We're gonna read some scary stories about, you know, parents that are tricked into wiring money to their, their kid who it sounds like is in, you know, some Mexican prison or something. But, you know, it turns out it was an attacker, you laugh but that that's gonna happen, happened recently and, and there's gonna be a bunch of those things and they're gonna make scary headlines. But I actually think it's going to accelerate us getting away from that idea that we can tell who to trust simply by how they sound or what they look like or those things. I think the longer term challenges for the CEO S that are in this room are gonna be much more around. How do they take advantages of what A I brings? But they do it in a way which is responsible over the long term. And so what we're starting to see is a lot of times employees at company are using A I to get their job done without understanding that the minute that they send that data off to an A I engine, it's gone for good, right? And it gets incorporated into it. A way of thinking it is imagine like a two year old and you accidentally swear in front of them. And then for the next period of time, all they're doing is just saying the bad word over and over and over again. A I is exactly like that you can't get the information back out of that. And that creates a real problem around trade secrets, that creates a real problem around data, locality and data compliance around the world right now, if you're using an A I system and you're a European based company, you are sending it all back to the United States and it's not even something where if you send a GDPR request and say, hey, forget that you can do it. It's embedded in a black box that is almost unforgettable. And so I think the thing that we're talking to our clients about is how do you make sure that as you're using A I, you're doing it anticipating that you wanna get those benefits, but at the same time, you wanna understand what the risks are and make sure that you're only letting the data flow out of your company. There's actually something you want to flow out of your company. Let's talk about the current state. Um I've been a leader of A I machine learning at some of the biggest companies in the world and A I, I've seen for the last decade drive truly digital transformation. Now, if it's the vehicle of trans of digital transformation, if anything generative A I is just speeding up that car on the highway and it's going super fast, what we are not doing are the basic maintenance, the basic safety and security checks on that vehicle. And so if I'm looking at where the risk is, the risk is really in the supply chain that we've seen, how are, how is A I being built? How are these models being built from the data, the foundational models, the open source libraries within the supply chain. So when we meet with customers, we ask two questions and I please take this back to your team. I ask how secure is your A I? And how do you know and how do you know is incredibly important because companies today don't have the capabilities or the ability to audit these systems and have visibility of their A I therefore, it's really hard to be able to secure it. Wow. I love that. Um Francois, you said that the number of cyberattacks on companies and applications have increased dramatically to the point where cyberattacks are costing companies $8 trillion a year globally. I found that number absurd, but in a crazy way, doesn't A I kind of arm the cyber criminals with tools that can now scale and what, what are do companies need to allocate more capital towards security? Well, yeah, the, the the cybercrime the cost, the cost around the world are around eight trillion, which is the third largest economy in the world after the US and China uh A I will accelerate that. Um And uh it, it does arm um cyber Attackers with more tools and therefore it will increase the speed and volume of attacks, not necessarily the sophistication that allows Attackers that would be not sophisticated at all to be able to launch an attack. One of the scary facts, for example is that today somebody that wants to attack, you would went somewhere doesn't actually know, doesn't actually need to know how to code that if you think about it. It is quite scary. Those folks can actually go online and you know, buy tools, there are tools like fraud GP T you can buy for $200 a month and gives you a starter kit for attacking companies with went somewhere and exploiting vulnerabilities immediately. Uh it doesn't mean that companies necessarily need to spend more money on, on cybersecurity. But it, it means that number one companies to be, need to be using the tools that exist today. I see a lot of companies that don't actually use the current generation of A I. There are a lot of tools today in security for security controls, for monitoring logs, for protecting against attacks that use the current generation of machine learning models to protect against these attacks. And too many times companies underestimate the sophisticated nature of their Attackers and therefore don't use those tools and get breached or under attack and don't even know it. And, and Ian uh what are some of the inherent risks uh That generative A I presents if those inputs in, in the safety belt isn't there uh when you're first starting. So a couple things as we think about generative A I and A I under the hood, it's, it's machine learning again. And when we talk about the differences and budgets that you brought up machine learning systems are different than all the other systems within your company. So as you're thinking about how do you protect your data, your infrastructure, your equipment, you need to think about how you're going to protect your A I. Now, what are some of those differences? The tools, the tools that developers and practitioners use to build models are different and today's security vendors are, don't scan those tools with a way to build its detect vulnerabilities. Also the models in the supply chain. Now, when you brought up generative A I, specifically, there's new types of attacks. One of them called prompt injection attacks. It's basically like a sequel injection attack where I can go in there and I can try to mislead the model to do something in a malicious way as a bad actor. Now, this becomes a lot more problematic when it's not just, you know, working with chat GP T to send me some information that you know, is crafting an email. It's when we start to connect these systems to agents, agents that carry out tasks. That's where the real risk lies. And therefore, that's why we need auditability, visibility and security of these systems. And Matthew, how are we seeing this practically uh at cloud flare? What measures are you taking to make sure that your company and the companies that you work with aren't victims of some sort of attack? Well, I think that the, the first thing and again to be the optimist on the panel is, you know, the whole thesis of cloud flare was that if we could get enough of the web behind us, if we could see enough information that we could feed that into what, what we frankly still are somewhat uncomfortable calling A I systems but are machine learning systems that could find new vulnerabilities and stop them before they were attacked so that you get the benefit of being part of the collective. That is, that is cloud flare. And what's been amazing in the same way that publicly we've all watched over the last 18 months as it seems like these A I systems went from jokes to, you know, writing sonnets, right? Which is pretty amazing. Um That same thing we've been watching internally where I remember about a year and a half ago, the first time our system found a vulnerability that no human had ever found before. And that went from something that was like, whoa, that's amazing to now, something that's happening on a very frequent and regular basis where we are oftentimes in front of these vulnerabilities and in fact, you know, have, have a number of those that were working with the industry in order to help clean up right now that we wouldn't have been able to find if it was simply humans that were out there. And so I think that the very value of this uh and the very value of using cloud services that you can uh have the benefits of all of the information, use that with the algorithms because the winners in the A I space aren't gonna, the algorithms are, are, are fairly, you know, trivial. They're, they're not, they're, they're the complexity, isn't there? The real thing that you need is to feed the algorithms with the right amount of data and in the security space every single day clouds or customers are safer because of the fact that being able to use these machine learning systems in order to protect everyone who uses our network. And, and again, that's where there are gonna be some scary stories that are out there. But I actually think that the good guys thankfully do have more resources than the bad guys. Um We, we have more access to data than, than they do. And there are ways that we're going to be able to not in the short term, short term is gonna be a little scary, but over the long term, I think the world becomes safer because of a I not, not scary. Have you ever gotten breached? Oh yeah, I mean, there's, we are constantly under attack. It's hard to imagine a threat actor that that isn't going after us. But what I really like is that the way that our systems are designed, we have an employee fall for a phishing scam, ok? And make sure that even when that happens, there's a belt and suspenders that keeps them from being protected and that is in part with machine learning system. So I'll give you a very specific example. I'm in Washington DC right now. Earlier this morning, I was in New York City. Um We have an A I system inside of cloud flow that is checking all of the our employees calendars to see where they are supposed to be and then automatically feeding rules back into the system where if I, um all of a sudden logged in right now to one of our systems, New York, it would simply, even if I had my user name and password and second factor authentication and used even my own machine because I've left it somewhere behind. That would still not let me through um our systems because we've been able to actually get smarter and more secure using the A I tools that we have internally and that we also then make available to all of our customers. Got it. Do we have any audience questions? No, everyone is sick of that. Um II I wanna ask Francois, how are you making sure that your workforce and the companies that you work with are up to speed and educated on protecting themselves from potential security risks? Well, so I think this is something every company out there needs to be doing, which is constantly training. Uh Our employees, Matthew mentioned earlier training, our employee against fishing and always tempting our, our employees with the fishing scam and see if they were doing that periodically or is it a one time training when they join the company? No, no, it's something that's on a constant basis. On a very regular basis. We train and retrain against these attacks. We analyzed employees who perhaps didn't get the training or fail for these attack and go back and, and train them on a very, very regular basis. That's something that we do today against regular attacks. But we're starting to train our employees on A I literacy because we feel that that's going to be something that every company needs to do to upscale their uh their employees and their knowledge to be able to both use the um uh A I in the most responsible way and understand best prac best practices for prompts. But also training uh engineers who are building these algorithms to know how to sanitize uh the data sets. Um And you know, be able to build these algorithms responsibly, you know, and I think that the thing though, that I would go even further on that is your employees are using A I, whether you want them to or not. And so the vast majority of employees are using A I and, and something like half are not telling their employers that they're actually doing it, they're using it for their, using it for their daily job to get to get work done. Because these are powerful algorithms. If you're having to write reports, you know, or you're having to generate code, it's a very safe place to or very productive place to be able to turn to what's again scary about it. It's a two year old. And so once you take your company secret and you send it to the two year old, it's gone forever, right? And that bad words gonna get said over and over and over again. And so I think that part of the tools here is not just in training, but actually putting in place a lot of the, you know, almost like D LP systems that say, you know, again, if you imagine there's a whole bunch of support documents, you're probably fine with those going to the A I systems and those learning. That's great because that actually is, is teaching the A is about how your product works. That's fine. You might have some other secrets that you say we can use on A I systems, but they have to be run locally inside of the company. And then there are other things, you know, that the, the sort of bad metaphor is the secret formula to Coca Cola. You may never want that to actually go into the, it's really tough because again, it's like the two year old, once these systems have learned something, it's not like a database row where you can go in and just delete it unlearning is actually this really fascinating area of A I research right now. And it's almost like a like AAA movie. It's like the, you know, turtle sunshine of the spotless mind. Like we don't have that system to have these, these A I systems unlearn things. And what that means is that it's actually a much bigger security threat for a company, even for insiders to make sure that as the data is leaving the systems, that it's the right data that's approved to be used. And that means training for sure, but it also means implementing the technical systems that restrict it and thinking about the supply chain, Ian, were you going to add something? Yeah, I think education is critical. I meet with 10 to 12 chief information security officers every week and I was meeting one with a big Pharma company this past week and he said, Ian, we're early in our A I journey. We're just now deploying A I. My response was no, you're not, you've been deploying A I for well over a decade as you look at drug discovery and every other facet of your business. But it's just now starting to get on to the table of the cyber security experts within the company. These systems again are different. We must educate how they're different and how we need to treat them differently in terms of looking for vulnerabilities, looking for exploits. And I would love to end on a very practical note. What are some tools that the business leaders here today can use to make sure that they're doing this responsibly and ethically other than promoting our own products? Um I think it's really just asking the simple questions and we heard it from Julie Sweet from accenture as well asking about who's in charge of responsible A I who's in charge of security of A I who's in charge of the visibility, the auditability of these systems just like your financial systems have a ledger. Your A I systems need to have a ledger. And oh, by the way, I only know five companies on the planet that I've worked with. I've worked with thousands that actually have that capability. That's where I would start. I, I'm not above promoting my own products, cloud players. A A I gateway gives you the tools to be able to control how data flows out of your systems. We're working with some of the largest companies in the world to make sure that they have control over that. We make it super easy for you to tag data and say here's where it goes and also control the budget that you have for your A I spend. Because the biggest thing that we hear is these budgets just blow up out of control in a way that even the cloud budget seem seem relatively limited. Awesome, great note to end on. Thank you to our panelists for joining us today.