The ART of Security Summit: Strategic Risk Management for CISOs | Rick Caccia, CEO, WitnessAI

Clips
More from The ART of Security Summit: Strategic Risk Management for CISOs

Rick Caccia

CEO

WitnessAI

play_circle_outline WitnessAI CEO Rick Caccia discusses AI security and risk reduction strategies.

play_circle_outline Importance of understanding and monitoring employee use of AI tools.

play_circle_outline Collaboration between security and other functions essential for effective AI integration.

Info

Rick Caccia, CEO, WitnessAI

Rick Caccia

CEO WitnessAI

search

Jackie McGuire

>> Welcome back to the ART of Security Summit. I'm Jackie McGuire, Practice Lead and Principal Analyst for Security at theCUBE Research. We're knee-deep in our avoid segment and I'm really happy to bring on one of my favorite thought leaders who is kind of threading a needle between avoid and reduce. Rick Caccia, he is CEO of WitnessAI and he's here to talk to us today about the most hot button topic I think we have, which is AI in AI security. Rick, welcome. Thank you for being here today.

Rick Caccia

>> Thanks, Jackie. It's awesome to be here. Thanks for having me.

Jackie McGuire

>> All right. So we just spoke to KnowBe4 and we were talking with KnowBe4 about training being your first line of defense. And when you and I were talking about how we can avoid and reduce risk associated with AI, one of the things that you had pointed out to me is that sometimes it's just about being aware, being aware of what's going on. So I was wondering if you could talk to us a little bit about what WitnessAI does, and then also about how you see education playing a part in security and AI in general.

Rick Caccia

>> Perfect. So let me give you a little blurb of when we started the company. One thing we learned and how it got us to where we are today. We started coming at this maybe the way most security startups do. You look at AI, you think there's going to be new types of attacks. You go to security officers, security departments and talk about this. And I'll tell you, the response we got for most of them was, "Yeah, I don't care. That's far off. I'm not going to treat this like you had another security attack." They said, "What do you care about?" I said, "What I really want to know is how are my people using this stuff? I want to see it. Because unlike past types of technologies, I want to enable AI. I don't want to block AI." So we give them the ability to get visibility, see what their employees are doing with first and third-party AI to put various policy controls in place and to make sure that their data and their people are protected. And then more recently, we've added analytics so they can understand how it's working. What does the risk look like? Which teams are using it the most? Which teams are using it the least? How do I optimize? And that's what we do.

Jackie McGuire

>> That's awesome. Yeah, I think it's a common refrain in security that you can't really protect what you can't see and you can't protect what you don't understand. So I think it's really important to emphasize that the first line of defense is even understanding what's going on in an organization. All right. So once you kind of have a feel for what's going on in your organization, what is the process that you generally see an enterprise go through as they go from, "Okay, what's going on?" to "Let's get a handle on it, let's create policies and let's enforce those policies"? How does that usually go?

Rick Caccia

>> Yeah. So the first thing is every single company we talk to has an idea in their mind of how their employees are or are not using this stuff. So it may be something like, "Hey, we block it all. Don't worry about it." And then you cut our product in, you find out that employees are bypassing all these things or there's AI apps they didn't know about. Or they say, "Hey, it's okay, we have the corporate version of ChatGPT, we have the corporate version of Copilot, don't worry about it." And then you find that there's all kinds of employees using personal version, free version, so forth, bypassed it. So the first thing is you have to give a picture of what they're actually doing. And the most interesting stat we find is you go into a customer and they think that their employees are using two or three AI apps. "Hey, they use ChatGPT, maybe Copilot." And then they find that they're using 20 and they freak out. But then you say, "Wait a minute. Is 20 bad? It's not necessarily bad. Maybe these few apps are powerful and the employees are getting their stuff done faster and better. Why would you want to stop this stuff?" They go, "Yeah, that's right. We should want to use it. We want to enable this use." And so then they shift the mindset to less how do we block and stop? And how do we more enable? How do we train? I think we had KnowBe4 on here before me, it's a great company, we integrate, we're in the middle of integrating with their product. And you could easily see how for a lot of these cases, "Hey, an employee's using something, you want to use a product like that to initiate some training. In other cases, you want to maybe block because their best intention doing something wrong. In other cases, you want to anonymize what they do and let it go through so they get answers." And it just sort of varies along that spectrum of risk, reward and productivity and security.

Jackie McGuire

>> And I think there's a saying about roads that are paved with the best intention. So generally, when people are using AI tools, it's not because they're trying to leak sensitive information or they're trying to get somebody's... And one of the things that I always think about with code is if you're not using AI tools that are specifically built for writing code, you could potentially pull in sources of protected code, things that are already owned by other people. So there's a lot of risk introduced, but also someone coming from data science who was not a software developer, didn't really understand Python. I wrote horrible code in the first several months that I was a data scientist. So I mean, it would've helped me a ton, but it's important to really put those protections in place. All right. So what happens... All right. Let's say, I have all these great detections in place to understand what my employees are doing. What happens if something goes awry? What happens if sensitive data goes somewhere it shouldn't? How do you detect that and how do you handle that?

Rick Caccia

>> Perfect. Yeah, I like your phrasing. I say AI is a matter of good intentions gone bad. And people, like you said, they're not trying to steal data, they're trying to get their jobs done. And take the example of code that you brought up. It's not just, "Hey, I'm writing Python." We're seeing things like in healthcare, we're seeing people use coding tools to write scripting around how to evaluate prior authorizations. It's a broad set there. You want to build... We built basically classification engines that look at the prompts being sent out and can classify them as, "Hey, this is code. What do you want to do?" And in some cases, you may say block. In some cases, you may say it's fine, let it go through. The more interesting thing we're seeing is customers just saying, "Well, I want my developers to use this stuff. I just don't want it to go out to GitHub." So we've trained an internal model on our code and I want you to intercept it and route it there. And then the best of both worlds, the developer gets an answer, but security's happy that the code never left the building. And if you think about DeepSeek, when DeepSeek hit the scene, there was some amazement and also concern because DeepSeek, the model itself wasn't a problem, but the APIs wherever hosted were the problem. So you may say, "Well, I want people to use a DeepSeek model, but only when it's running on my infrastructure. So I know the APIs are secure. Can you intercept and route?" And that's not a DLP problem. It's a very different, "How do I use AI productively?" kind of problem. To make it work, you have to understand the network, you have to understand the content, and you have to understand how you want your policies to work.

Jackie McGuire

>> And I guess, I know we both know that every organization has done a really good job with data classification and identification and categorization, and it's all stored in a really efficiently indexed way. But let's say an enterprise just happens to be one of the exceptions that has dirty data. Do you have a way to intelligently identify what data is? And other than as a data scientist, several years ago, I used Regex, like pattern matching, a very inefficient way to do things and very error-prone. Is there a better way to do it? And what have you found that to be?

Rick Caccia

>> Yeah. So here's the very interesting thing. So tagging data and scanning data in flight, DLP, data loss prevention, has been around for 25 years. And it's one of those markets that if I'm being charitable, I would say customers have mixed feelings about how effective it is. And so when we looked at this, we said, "Well, we can integrate with these DLP products and classify it, but we've done something that's pretty unique. We have the ability to look at the prompt and not necessarily decide if the data itself is confidential, but it's someone trying to do something they shouldn't do. So intention based routing." So we could say, "Hey, Jackie is working with a data set and maybe I'm going to run that through a DLP product and decide if it's confidential." But I can also say, "Hey, this looks like, I don't know, is she asking to write a contract or a marketing analysis?" And I've put a policy in place that says, "Well, only the people in data science can do that. Jackie's not in data science, so this intention, write this analysis, I'm going to reroute it or I'm going to block it." So it's not just looking at the data itself, it's looking at the intentions and taking action there. And you can sort of Hail Mary, throw the football right to the end zone overall the defenders and say, "I'm not going to worry about the classification, I'm going to add another way to block at the intention level regardless of what the data is." And we think when companies come up with their acceptable use policies, identity and intention together are a really good way to control data loss and harmful use.

Jackie McGuire

>> Yeah. I definitely think that logging and data use policies are where we're going. One of the things that's always surprised me in every area of corporate America that I've consulted in is lack of policy around things. And I think it starts with something's too new. "Oh, we don't really know how we're going to use it, so we can't really write a policy." And then that morphs into, "Well, we've never had a policy. And we kind of know how people use it." I don't think that's going to fly given kind of the nature of how quickly things can spiral out of control with the exponent of compute we use in AI. So I think what I'm taking is the TLDR is first understand what's going on in your environment, understand what the tools are, how they're being used, and whether they really do pose a threat or whether they can have some reasonable kind of fencing put around them. Once you've done that, then you can kind of start looking at how people are using them individually and what their intentions are with the data that they're using and put in place some protections there that involve the right people. I think when we were talking about this before we started recording, I was telling you that I saw this study that said that less than 30% of AI projects have a security person involved, anybody from the security function before POC, which is a little bit terrifying. And I guess, if you're a small bootstrapped organization or you're trying to move fast, maybe the security team doesn't know when people spin up new databases, they don't know when people spin up new buckets. So I think if you could understand more about how that's happening and then put some reasonable protections around it. And then is there anything I'm missing? Is there a final step in the process that-

Rick Caccia

>> No, I think, that's a really good way to put it, Jackie. So AI is different though. I believe it's different in that even security organizations want to become the department of yes. So the path to getting there is first understand, first understand how it's being used, then you can decide, "Okay, how do we shape our policies?" And I think you brought up a really good point around security being involved in the minority of these things. I think in our experience, security shows up at the tail end when the marketing department, the data science team, whatever, shows up and says, "Hey, security department, bless this new app we've built." And security has no idea it was being built. I just talked to a CISO earlier today in a sales column. He said, "Boy, it would be great if they came to me at any point in the process. Usually, I find out after it's been launched." So that observability and visibility is a big deal. And then I think the second big myth is that companies have these policies in place. Most don't, they're trying to figure it out. So how do I take a legal policy document for acceptable use and map it to AI use for employees and customers? You got to see what you've got there before you make the policies and then you have to figure out how to implement them.

Jackie McGuire

>> Yeah. Security policy in general is tough because it's one of the most, there are no two security stacks that are identical. There just aren't. And there are a few other areas in the organization where you run into, I'm still dealing with technology that's literally 30 years old and trying to make it work along with technology that we just bought and are rolling out. So it'll be really interesting to see whether security moves into other functions and they have overlays, because everything's a security issue now. And the more you involve data, the more you involve AI, the more it's a security problem. So I'll be interested to see whether enterprises choose to make their non-security functions more security literate and aware, or whether we're going to have to have embedded security functions within a lot of different functions that we didn't use to so that we can start earlier in the process. Because I doubt the threats going away or going to decrease, and I don't think we're going to be able to have AI fight AI for a while.

Rick Caccia

>> I agree. Look, here's, to me, the closest analogy to this. The first startup I worked at was in the dot-com time and web application servers. And I think there was a period where there were web sites and everything else that wasn't web. And the notion today of having applications that aren't internet enabled is just ludicrous. And I think the same thing's going to happen with AI. So you're talking about security sort of melting into other functions. I think you've got to do that because AI is going to blend into everything else. And just like you wouldn't imagine having an app that doesn't have web access, in a couple of years, we won't have anything that doesn't have AI embedded in it. So being able to understand it, make sense of it, use other AI to figure it out, not to stop attacks, but to figure out what's going on. And how do I handle this? Is it safe? Is it not safe? Do I reroute it? Do I warn? Do I train? You're going to see that become an imperative, I think, at most companies pretty soon.

Jackie McGuire

>> Well, I feel like you stole my job, because I can't imagine a better way to wrap it up than what you just said. And I'm probably going to steal that soundbite for a clip because I think you just summarized it really well. We need to embrace this. We need to be prescriptive about what kinds of AI, LLMs should be used for which things and which more efficient machine learning and AI. But it's going to be here, it's already here. So I think the best thing we can do is embrace it so that we become the department of how do we get to yes, not no, just not right now. This is why we can't say yes, but we can get to yes and really start to take advantage of some of the good sides of AI, not just the risks. All right. So thank you so much for being here today, Rick. This was amazing. I'm so glad we were able to get you on. Witness is able to help avoid and reduce risk, which is amazing to me. And I was really excited about the product when we talked about it. So hopefully people can check it out and be more AI-friendly. It's something that I think security is terminally snarky anyway, but I think that attitudes are shifting. So thank you for being here to help us have a more open attitude towards ways we can put boundaries in place that aren't obstructive.

Rick Caccia

>> Thanks, Jackie. I appreciate being here. It was great talk. Thanks a lot.

Jackie McGuire

>> Thank you everybody for the ART of Security Summit. I'm Jackie McGuire. We will be back shortly with more tips on how to avoid, reduce and transfer risk.

Rick Caccia, CEO, WitnessAI

search

Jackie McGuire

>> Welcome back to the ART of Security Summit. I'm Jackie McGuire, Practice Lead and Principal Analyst for Security at theCUBE Research. We're knee-deep in our avoid segment and I'm really happy to bring on one of my favorite thought leaders who is kind of threading a needle between avoid and reduce. Rick Caccia, he is CEO of WitnessAI and he's here to talk to us today about the most hot button topic I think we have, which is AI in AI security. Rick, welcome. Thank you for being here today.

Rick Caccia

>> Thanks, Jackie. It's awesome to be here. Thanks for having me.

Jackie McGuire

>> All right. So we just spoke to KnowBe4 and we were talking with KnowBe4 about training being your first line of defense. And when you and I were talking about how we can avoid and reduce risk associated with AI, one of the things that you had pointed out to me is that sometimes it's just about being aware, being aware of what's going on. So I was wondering if you could talk to us a little bit about what WitnessAI does, and then also about how you see education playing a part in security and AI in general.

Rick Caccia

>> Perfect. So let me give you a little blurb of when we started the company. One thing we learned and how it got us to where we are today. We started coming at this maybe the way most security startups do. You look at AI, you think there's going to be new types of attacks. You go to security officers, security departments and talk about this. And I'll tell you, the response we got for most of them was, "Yeah, I don't care. That's far off. I'm not going to treat this like you had another security attack." They said, "What do you care about?" I said, "What I really want to know is how are my people using this stuff? I want to see it. Because unlike past types of technologies, I want to enable AI. I don't want to block AI." So we give them the ability to get visibility, see what their employees are doing with first and third-party AI to put various policy controls in place and to make sure that their data and their people are protected. And then more recently, we've added analytics so they can understand how it's working. What does the risk look like? Which teams are using it the most? Which teams are using it the least? How do I optimize? And that's what we do.

Jackie McGuire

>> That's awesome. Yeah, I think it's a common refrain in security that you can't really protect what you can't see and you can't protect what you don't understand. So I think it's really important to emphasize that the first line of defense is even understanding what's going on in an organization. All right. So once you kind of have a feel for what's going on in your organization, what is the process that you generally see an enterprise go through as they go from, "Okay, what's going on?" to "Let's get a handle on it, let's create policies and let's enforce those policies"? How does that usually go?

Rick Caccia

>> Yeah. So the first thing is every single company we talk to has an idea in their mind of how their employees are or are not using this stuff. So it may be something like, "Hey, we block it all. Don't worry about it." And then you cut our product in, you find out that employees are bypassing all these things or there's AI apps they didn't know about. Or they say, "Hey, it's okay, we have the corporate version of ChatGPT, we have the corporate version of Copilot, don't worry about it." And then you find that there's all kinds of employees using personal version, free version, so forth, bypassed it. So the first thing is you have to give a picture of what they're actually doing. And the most interesting stat we find is you go into a customer and they think that their employees are using two or three AI apps. "Hey, they use ChatGPT, maybe Copilot." And then they find that they're using 20 and they freak out. But then you say, "Wait a minute. Is 20 bad? It's not necessarily bad. Maybe these few apps are powerful and the employees are getting their stuff done faster and better. Why would you want to stop this stuff?" They go, "Yeah, that's right. We should want to use it. We want to enable this use." And so then they shift the mindset to less how do we block and stop? And how do we more enable? How do we train? I think we had KnowBe4 on here before me, it's a great company, we integrate, we're in the middle of integrating with their product. And you could easily see how for a lot of these cases, "Hey, an employee's using something, you want to use a product like that to initiate some training. In other cases, you want to maybe block because their best intention doing something wrong. In other cases, you want to anonymize what they do and let it go through so they get answers." And it just sort of varies along that spectrum of risk, reward and productivity and security.

Jackie McGuire

>> And I think there's a saying about roads that are paved with the best intention. So generally, when people are using AI tools, it's not because they're trying to leak sensitive information or they're trying to get somebody's... And one of the things that I always think about with code is if you're not using AI tools that are specifically built for writing code, you could potentially pull in sources of protected code, things that are already owned by other people. So there's a lot of risk introduced, but also someone coming from data science who was not a software developer, didn't really understand Python. I wrote horrible code in the first several months that I was a data scientist. So I mean, it would've helped me a ton, but it's important to really put those protections in place. All right. So what happens... All right. Let's say, I have all these great detections in place to understand what my employees are doing. What happens if something goes awry? What happens if sensitive data goes somewhere it shouldn't? How do you detect that and how do you handle that?

Rick Caccia

>> Perfect. Yeah, I like your phrasing. I say AI is a matter of good intentions gone bad. And people, like you said, they're not trying to steal data, they're trying to get their jobs done. And take the example of code that you brought up. It's not just, "Hey, I'm writing Python." We're seeing things like in healthcare, we're seeing people use coding tools to write scripting around how to evaluate prior authorizations. It's a broad set there. You want to build... We built basically classification engines that look at the prompts being sent out and can classify them as, "Hey, this is code. What do you want to do?" And in some cases, you may say block. In some cases, you may say it's fine, let it go through. The more interesting thing we're seeing is customers just saying, "Well, I want my developers to use this stuff. I just don't want it to go out to GitHub." So we've trained an internal model on our code and I want you to intercept it and route it there. And then the best of both worlds, the developer gets an answer, but security's happy that the code never left the building. And if you think about DeepSeek, when DeepSeek hit the scene, there was some amazement and also concern because DeepSeek, the model itself wasn't a problem, but the APIs wherever hosted were the problem. So you may say, "Well, I want people to use a DeepSeek model, but only when it's running on my infrastructure. So I know the APIs are secure. Can you intercept and route?" And that's not a DLP problem. It's a very different, "How do I use AI productively?" kind of problem. To make it work, you have to understand the network, you have to understand the content, and you have to understand how you want your policies to work.

Jackie McGuire

>> And I guess, I know we both know that every organization has done a really good job with data classification and identification and categorization, and it's all stored in a really efficiently indexed way. But let's say an enterprise just happens to be one of the exceptions that has dirty data. Do you have a way to intelligently identify what data is? And other than as a data scientist, several years ago, I used Regex, like pattern matching, a very inefficient way to do things and very error-prone. Is there a better way to do it? And what have you found that to be?

Rick Caccia

>> Yeah. So here's the very interesting thing. So tagging data and scanning data in flight, DLP, data loss prevention, has been around for 25 years. And it's one of those markets that if I'm being charitable, I would say customers have mixed feelings about how effective it is. And so when we looked at this, we said, "Well, we can integrate with these DLP products and classify it, but we've done something that's pretty unique. We have the ability to look at the prompt and not necessarily decide if the data itself is confidential, but it's someone trying to do something they shouldn't do. So intention based routing." So we could say, "Hey, Jackie is working with a data set and maybe I'm going to run that through a DLP product and decide if it's confidential." But I can also say, "Hey, this looks like, I don't know, is she asking to write a contract or a marketing analysis?" And I've put a policy in place that says, "Well, only the people in data science can do that. Jackie's not in data science, so this intention, write this analysis, I'm going to reroute it or I'm going to block it." So it's not just looking at the data itself, it's looking at the intentions and taking action there. And you can sort of Hail Mary, throw the football right to the end zone overall the defenders and say, "I'm not going to worry about the classification, I'm going to add another way to block at the intention level regardless of what the data is." And we think when companies come up with their acceptable use policies, identity and intention together are a really good way to control data loss and harmful use.

Jackie McGuire

>> Yeah. I definitely think that logging and data use policies are where we're going. One of the things that's always surprised me in every area of corporate America that I've consulted in is lack of policy around things. And I think it starts with something's too new. "Oh, we don't really know how we're going to use it, so we can't really write a policy." And then that morphs into, "Well, we've never had a policy. And we kind of know how people use it." I don't think that's going to fly given kind of the nature of how quickly things can spiral out of control with the exponent of compute we use in AI. So I think what I'm taking is the TLDR is first understand what's going on in your environment, understand what the tools are, how they're being used, and whether they really do pose a threat or whether they can have some reasonable kind of fencing put around them. Once you've done that, then you can kind of start looking at how people are using them individually and what their intentions are with the data that they're using and put in place some protections there that involve the right people. I think when we were talking about this before we started recording, I was telling you that I saw this study that said that less than 30% of AI projects have a security person involved, anybody from the security function before POC, which is a little bit terrifying. And I guess, if you're a small bootstrapped organization or you're trying to move fast, maybe the security team doesn't know when people spin up new databases, they don't know when people spin up new buckets. So I think if you could understand more about how that's happening and then put some reasonable protections around it. And then is there anything I'm missing? Is there a final step in the process that-

Rick Caccia

>> No, I think, that's a really good way to put it, Jackie. So AI is different though. I believe it's different in that even security organizations want to become the department of yes. So the path to getting there is first understand, first understand how it's being used, then you can decide, "Okay, how do we shape our policies?" And I think you brought up a really good point around security being involved in the minority of these things. I think in our experience, security shows up at the tail end when the marketing department, the data science team, whatever, shows up and says, "Hey, security department, bless this new app we've built." And security has no idea it was being built. I just talked to a CISO earlier today in a sales column. He said, "Boy, it would be great if they came to me at any point in the process. Usually, I find out after it's been launched." So that observability and visibility is a big deal. And then I think the second big myth is that companies have these policies in place. Most don't, they're trying to figure it out. So how do I take a legal policy document for acceptable use and map it to AI use for employees and customers? You got to see what you've got there before you make the policies and then you have to figure out how to implement them.

Jackie McGuire

>> Yeah. Security policy in general is tough because it's one of the most, there are no two security stacks that are identical. There just aren't. And there are a few other areas in the organization where you run into, I'm still dealing with technology that's literally 30 years old and trying to make it work along with technology that we just bought and are rolling out. So it'll be really interesting to see whether security moves into other functions and they have overlays, because everything's a security issue now. And the more you involve data, the more you involve AI, the more it's a security problem. So I'll be interested to see whether enterprises choose to make their non-security functions more security literate and aware, or whether we're going to have to have embedded security functions within a lot of different functions that we didn't use to so that we can start earlier in the process. Because I doubt the threats going away or going to decrease, and I don't think we're going to be able to have AI fight AI for a while.

Rick Caccia

>> I agree. Look, here's, to me, the closest analogy to this. The first startup I worked at was in the dot-com time and web application servers. And I think there was a period where there were web sites and everything else that wasn't web. And the notion today of having applications that aren't internet enabled is just ludicrous. And I think the same thing's going to happen with AI. So you're talking about security sort of melting into other functions. I think you've got to do that because AI is going to blend into everything else. And just like you wouldn't imagine having an app that doesn't have web access, in a couple of years, we won't have anything that doesn't have AI embedded in it. So being able to understand it, make sense of it, use other AI to figure it out, not to stop attacks, but to figure out what's going on. And how do I handle this? Is it safe? Is it not safe? Do I reroute it? Do I warn? Do I train? You're going to see that become an imperative, I think, at most companies pretty soon.

Jackie McGuire

>> Well, I feel like you stole my job, because I can't imagine a better way to wrap it up than what you just said. And I'm probably going to steal that soundbite for a clip because I think you just summarized it really well. We need to embrace this. We need to be prescriptive about what kinds of AI, LLMs should be used for which things and which more efficient machine learning and AI. But it's going to be here, it's already here. So I think the best thing we can do is embrace it so that we become the department of how do we get to yes, not no, just not right now. This is why we can't say yes, but we can get to yes and really start to take advantage of some of the good sides of AI, not just the risks. All right. So thank you so much for being here today, Rick. This was amazing. I'm so glad we were able to get you on. Witness is able to help avoid and reduce risk, which is amazing to me. And I was really excited about the product when we talked about it. So hopefully people can check it out and be more AI-friendly. It's something that I think security is terminally snarky anyway, but I think that attitudes are shifting. So thank you for being here to help us have a more open attitude towards ways we can put boundaries in place that aren't obstructive.

Rick Caccia

>> Thanks, Jackie. I appreciate being here. It was great talk. Thanks a lot.

Jackie McGuire

>> Thank you everybody for the ART of Security Summit. I'm Jackie McGuire. We will be back shortly with more tips on how to avoid, reduce and transfer risk.