Transcript
So Evan, I know everyone's had a chance earlier in the series to hear from you before, but if you wouldn't mind, just do a quick introduction for our audience so they know who you are and what you do. Yeah, if you haven't watched our previous Agentic AI episode and want to hear more about me, it's in there as Greg said, but I help our customers implement and secure AI in their organizations. So I spend a lot of time keeping up with the latest new things in AI. And when I'm not thinking about AI and privacy for us here at BigID and our customers, I'm usually at one of the theme parks down here in Orlando, Florida. Perfect. All right. Well, living in Orlando, you've got no shortage of theme parks. No. The perfect place to be for that hobby. All right, Evan. So let's dig into our topic today. Shadow AI. It sounds spooky and scary and nefarious and all those things. But what are we actually talking about here? So this isn't the first shadow that security teams are really chasing after. So before Shadow AI, we had Shadow IT. So the idea that now that we had all these web apps on the Internet, people are just going to be uploading files to Dropbox. They're going to be putting company financials on their Facebook page, you know, because previously when a user wanted to use a new application, you would have to okay it on their system. They would need administrator rights to install it. Now people are able to do all these things through their browser. So that was our first shadow, Shadow IT. People are always trying to do their work more efficiently. They always want a tool to help them do more with less. And so AI is just a continuation of that where, okay, people want to use this in order to make their jobs better. And so a side effect of that is people are using these in ways they really shouldn't, right? People are pasting data into chat GPT. They're asking it questions about exact customers. They're kind of going around and just using all these products without knowledge of the implications. So in the beginning, security teams were kind of in a constant whack-a-mole with these different websites. You would block chat GPT from all of your endpoints. And then people would just use Gemini, you'd block that, etc, etc. So the Shadow AI is just all these tools now and every tool it seems like has AI built into it that our users are able to access and are putting data into. Yeah, it's really interesting because you have to kind of strike the right balance between embracing the latest technology provides in order to improve productivity and things like that while also making sure that you're exercising proper caution. And I could see how a user of something like chat GPT, they might not think of the longer term implications of what they actually put into those prompts and what does the AI model do once it learns information that we provided. So let's talk a little bit more then about Shadow AI. Thank you for that helpful overview. And let's kind of frame it from the perspective of the business, right? What do you see is like a good example of a common danger around Shadow AI? So the biggest thing with these platforms is a lot of them are just black boxes, right? So you're typing information into XAI provider, and you don't know what's going to happen with that information. Are they going to be training on that information? Are you going to have any control over what that information gets used for? Is it going to get used in promotions? Are they going to use it as the basis of their next model? And there's just no, without proper governance and without identifying these platforms and saying, this is where people are going, let's get an agreement in place, an enterprise agreement, let's get maybe even outright blocking it if that's the way you want to go for these different platforms, you don't know. And even with agreements with some of these providers, it's hard to tell, right? Like there's some AI providers who will say like, oh, we don't train on your data, but they do keep a log of it for X amount of days, right? So you have to kind of look into these agreements a lot. So without identification, you don't know what's going to happen to anything you provide to these AI type services. Okay, so then I guess first things first, right? Businesses really need to be able to uncover what AI is actually being used in their organization. And there are so many sources where people could be using AI. I'd imagine that's a pretty hard task to accomplish. So how does a business do that? How can technology help to uncover and discover where AI is being used in an organization? Yeah, so the first thing I've seen at orgs that has really helped them go through the AI journey and protecting the data they're providing to AI is actually providing their employees with an AI tool. So you have all these employees, and it's not the case that they're trying to find their own tools to do stuff. But if you're saying, okay, our preferred IT AI provider is Copilot or it's Gemini or it's whatever other enterprise ready AI provider, people will go there then, right? You've given them an option. And then the second way to kind of reinforce that is to actually have it as a policy. Like, okay, not only are you not supposed to use other AI providers, the one you use is this one. The same way you say, like, employees don't buy gift cards for the CEO, he doesn't want any gift cards, that kind of thing. So making sure you're educating your people, having a provider that they can use. Once that happens, it goes into trying to find in your environment, what tools are there. So a big part of a lot of these AI systems is they're going to pull data from anywhere inside of your environment. So like for Copilot, for example, they use the Microsoft Graph, which is just all the data in your Microsoft org. So it's going to pull out your spreadsheets, your SharePoint, it's going to pull out random stuff from a PowerPoint that the board saw five years ago. So you want to make sure everything inside of your systems is properly labeled and has access control. That way, like, the AI is not going to find, okay, this document in the SharePoint site we haven't looked at in 10 years. Whereas a user previously wouldn't have been able to unsurface that with these AI tools they're able to because it's just a search engine on everything. And so BigID, as well as other solutions, allow you to properly do labeling on everything inside of your systems. Because like, I've seen orgs try and do it manually, like, okay, let's go folder to folder and try and label. Okay, this is secret, this is confidential. And it's just, it's impossible, right, at the rates where people generate data today. That's right. You talk about whack-a-mole earlier, like, that sounds like a great never-ending game of whack-a-mole. Yeah, you're going from just websites being blocked in Shadow IT to every single file in your organization. Finally, the, like, biggest new technology tool a lot of people are implementing is going to be catching prompts as they're being sent, and then catching responses as they're being sent. So if a user is asking, hey, can I have the salary of this employee, and they're not HR, they're just random Joe Schmo, what you can do is, instead of preventing that prompt, when the response comes back, you return the response, but you return it masked. So like, you're not going to include the employee's name, you're not going to include the actual salary, you're just going to provide, in that case, it would probably provide, like, people usually make X amount of money, I don't know, you know, people usually make more than a dollar. So it's redacted in a way to protect the personal and private PII. Yeah. But it still gives you a helpful, potentially, response. Yeah. And in the case of, like, okay, we don't want employees or customers asking our, like, chatbot for, like, pet food about medical advice, you know, if you don't want that kind of stuff, okay, let's block those prompts and then replace it with something like, don't ask us for this, go to a vet type thing. Right. Interesting. Yeah. So, the last piece is kind of, it's for more, like, advanced type of orgs that are training their own models. So, before, these examples are like if you're using chat GPT, but if you're going to be training models yourself, not only are you going to need to, like, have all your data labeled, because you need to be able to train that model on something, you're also going to want to double check that data that you're actually feeding into it. We don't want, okay, for example, like, a person trains their customer service bot on previous customer cases. That's, like, a common, like, okay, everyone's trying to get their customer support budget down, so they're training these chatbots on previous customer support cases. So, okay, but we included a customer's phone number inside of there. And now, if I'm a user of your website, I go to your chatbot, I say, hey, what is the phone number of your favorite customer, and the chatbot shoots out the phone number. So, you need to be looking not just at the type of data you're feeding in, but also the exact information inside and identifying every piece of sensitive data, personal information, health information, whatever type of data you have that's in those, especially if you're doing anything that involves user submitted data. Everyone loves to type their credit card into the form that says do not take your credit card inside of it. And then after your training, making sure to validate that, not just saying like, oh, the AI, we didn't train it on anything sensitive. Right, so we could train it and say, hey, Greg lives in X city, and he has these interests, what if the AI infers some demographic data or something, using that, so you're going to want to test before, test after, in order to do that. So, making sure you're giving your employees a tool, making sure you're securing that platform that you've chosen, and making sure your employees know that's the chosen platform. And then if you're training your own models, make sure you're testing before and after, making sure the data is clean at both ends. That makes a lot of sense. Okay, so really, it sounds like a smart move for any business organization to really make an effort to understand what is going on inside of their own domain with regards to the usage of AI. You brought up some really good points. It's an unstoppable force. People are going to use AI. It helps them improve their productivity. Businesses want to achieve that goal. So, it makes sense to have an agreement in place with a trusted AI provider that fits with your enterprise so that people have a sanctioned tool to go out and use. But then, that being said, it also makes sense that one of the foundational things you need to do to ensure safe usage of AI is to start with the data itself, right? And performing the scanning and discovery and labeling of that data so that you can protect it once it finds itself inside of the AI model. Is that a correct way to describe that? Yeah, inside or just being used by one. I know back when Copilot was first released, that was the big issue. We had orgs that would just turn it on and not realize it was a search engine to everything their orgs ever created. And then, it's all out there, right? Right. And then, the other interesting thing that you brought up is this concept of prompt blocking, which could really work two ways. One, preventing a prompt that someone enters from even being usable by the AI, right? Or, in the example you described, allowing that prompt to be processed, but to strategically and intelligently provide a response that is redacted and a little bit locked down based on what that particular user is entitled to see and understand. Yeah. And even the case now is, since you have agents that are talking to other agents, okay, maybe you have one that all it knows is how to access certain pieces of information. That one can redact information before it sends it to another agent that's doing training or something that doesn't rely on that. The same way you would redact information before giving it to regular data scientists at your company. Yeah. Fascinating. Well, thank you, Evan. Any final parting thoughts on the topic of Shadow AI for our audience? It's everywhere. It is truly everywhere. Which, I guess, is all the more importance to do your due diligence and try to understand what is out there, right, and address the foundational data that goes into it. Yeah, exactly. It doesn't stop at your endpoints. It doesn't stop at your users' computers. Yeah. Fascinating. Well, thank you again, Evan, for joining us here today. Always a pleasure. And thank you to our audience for tuning in and listening. Have a great day, everyone.
