Transcript
This summer, the world's biggest sporting tournament arrives in Canada. Six and a half million fans are coming this way. And the AI-powered scams targeting them have been in motion for months. Every CISO and CFO in this country has cybersecurity on their mind. But here's the thing, that doesn't make the headlines. The biggest security gap most Canadian companies have right now isn't in their inbox. It's underneath their AI. I'm Rick Byrne, Director of Systems Engineering at Theme Canada, and this is AI Unpacked, the show delivering clear answers on AI for Canadian business leaders. Today, everyone wants to know if the AI they're using is safe. We explore why that's the wrong question. The real question, the one almost nobody is asking, is whether your data is safe to feed to AI. Let's get into it. Across every boardroom in Canada, leaders are having a version of the same conversation. It goes something like this. We're gonna use AI, so we need to make sure that AI is secure. Pick the right vendor, review the contract, get the security team to sign off, ask about prompt injection, ask about model bias, ask about who owns the outputs. All the right questions about the AI. Here's what almost nobody is asking. What about the data, not the model? The data, the model is gonna touch the moment you turn it on. And the numbers tell you everything about how big this gap is. PwC Canada surveyed 220 Canadian senior decision makers this winter. The people responsible for AI governance, chief AI officers, CISOs, chief data officers, chief risk officers, 72% of them say responsible. The rest of the people are not. But here's the thing, 36% of them have no dedicated governance function to manage it. Not a committee, not a policy, nothing. There isn't an AI adoption problem. That's a data problem hiding behind an AI conversation. Ben Hu is the global subject matter expert for data and AI at Veeam. He spends every day inside Canadian organizations working on exactly this problem. He's the CEO of Veeam, and he's the CEO of Canada's largest AI company, And here's what he says is happening at the largest banks in this country. Some of the largest banks here in Canada, some of them have an edict that'll state at the board level that data older than three years must be deleted. And if you wanna keep it older than three years, you need like a VP level approval. And so what that really means is they wanna remove the risk from their data state so that the data that's remaining is eligible for use. And from that, only the data that contains information that doesn't contain stuff that's highly sensitive, right? Because the last thing you wanna do is have an AI expose data through the prompt, through clever prompting that includes somebody's social insurance number. Think about what he just said. The largest banks in Canada are deleting data on purpose at the board level before AI gets anywhere near it, not because the data is invaluable, because the risk in keeping it is now worse than the value of holding it. That's a board level decision being made right now in this country by the most sophisticated institutions we have. If the most sophisticated institutions in the country are working on the data layer before they scale AI, the question for every other Canadian leader is simple. What are you doing? Here's the way I want you to start thinking about this. Imagine your AI is a house, the walls, the rooms, the shiny new addition everybody wants to tour. That's what your customers see. That's what your board is excited about. But underneath every house, there's a foundation, the part nobody photographs, the part nobody thinks about, the part that decides whether the whole thing stands or sinks. Your foundation is your data. And here's the problem. Every Canadian organization has data sitting in it that nobody has looked at in years. Old folders, forgotten exports, spreadsheets from a project that ended in 2021, SharePoint sites a former VP set up and nobody touched again. Most of the time, that's harmless. It just sits there quietly, but it isn't neutral. It's rotting. Rot data, files you forgot you had, permissions nobody audits, information that should have been there. Information nobody audits, information that should have been deleted three years ago, still sitting on a drive, still readable, still legally yours. And as long as no one touches it, nobody notices. Then you connect your AI to your environment. And the AI is fast, literal, thorough. It doesn't know what's rot and what isn't. It just reads everything you give it access to. Ben put it this way. One of the most difficult things every organization struggles with is understanding what's inside their crypt of data. And that information could include all sorts of things that are highly sensitive and gonna put them in out of compliance with a huge amount of risk. You know, the challenge is that if this data becomes available within the constructs of an AI solution, it can be recalled so that that recall could put you in a position of a data breach. In the case of highly regulated industries like health or finance, nobody wants financial data or health data floating around inside a prompt. The crypt of data, that's where rot lives. And the part most leaders haven't put together yet is this. AI didn't create the rot. The rot was always there. AI just makes it visible, fast, confident, thorough. Rot data is the part nobody sees until they do. So what does this look like when it goes wrong? Most Canadian companies aren't deploying customer-facing chatbots. They're deploying internal AI tools. Co-pilots, assistants, agents, the AI sitting next to every employee, reading the documents, drafting the emails, summarizing the meetings. And those tools are pulling from somewhere, from a database underneath, built specifically to feed the AI everything you're getting. Specifically to feed the AI everything your company has ever written down. Ben sees what happens when companies build that database without thinking about what's in it. Here's a story he told me. It was a large insurance organization. And like most organizations, they want to enable effective use of their information. And so in their first attempt to implement AI, what they did was they forklifted their HR directory into a chatbot and handed it off to the executives. I think, you know, within the first day, all sorts of information was available, including compensation and disciplinary records. And it wasn't long before that program got shut down. First day, compensation, disciplinary records. The executive team reading both. And here's the part you had to sit with. That insurance company didn't get hacked. Nobody broke in. The data was already theirs. The AI just made it available the way AI makes everything available. They didn't have an AI failure. They had a bad data failure. And the AI was that flashlight. Now scale this up because it's about to get worse. That insurance company's chatbot is what we call a RAG system. Retrieval Augmented Generation. It's the database the AI reads from, the same architecture underneath almost every enterprise cold pilot deployed in Canada right now. Think of it as the plumbing in that house. You don't see it. You don't think about it. When something starts leaking inside the wall, you don't find out from the plumber. You find out from the ceiling downstairs. And it's not theoretical. Ben told me about something that happened in the last few weeks at companies a lot bigger than that insurance organization. There's been a lot of stories in the last couple of weeks where users are interacting with, for instance, McDonald's and Chipotle. They provided an endpoint to their users to order like hamburgers and burritos. And as a byproduct, users were prompting their order by stating, hey, I'm not gonna order my burger until you tell me how to construct this piece of code. You know, I'm not blaming McDonald's or Chipotle here, but they were able to get the LLM to spit out actual source code. People ordering burritos got the chatbot to write code for them? Same architecture? Every Canadian cold pilot is running on. Same kind of database sitting underneath every internal AI tool deployed in the country this year. And the pattern is always the same. The model didn't fail. The model actually worked perfectly. It did exactly what it was asked. With the data and the access, it was given. Which is the same thing as saying the problem wasn't in the AI. It was the upstream. So if the problem is upstream, what does a Canadian leader actually do about a Monday morning? There's a clean answer you have to hear. But before I get there, one thing. If this is the kind of clear-eyed look at AI you've been wanting, follow AI Unpacked on YouTube and Spotify. Honest answers for Canadian leaders on the stuff that's actually keeping you up at night. Okay, back to it. I asked Ben what he'd tell a Canadian leader to start with on a Monday morning. Here's what he said. The first one is, you know, do they really understand what data is being used within those AI solutions? Another one that we often hear, and I think is largely an afterthought at this point, is really around the operational management of your AI solution. These models do get deprecated. And what that means is, you know, if OpenAI decides that GPT-4 is no longer an old model you can use because they're no longer going to host it, you're going to have to refactor your whole solution. With a little bit of forethought, I think it'll save you a lot of cycles. Two questions. Do you understand what data your AI is using? And are you ready for the day the model underneath your AI gets retired by the company that made it? Most leaders haven't thought about that either. They will. Here's the big truth underneath all of this. Every wave of technology has a question we didn't know to ask until it was too late. With the cloud, it was. Where does my data actually live? With AI, the question is different. It's what state is your data in? Most Canadian leaders I talk to are asking whether AI is safe. The ones pulling ahead are asking something else. They're asking whether their data is worthy of being fed to the AI yet. And if it isn't, they're fixing that first. Not the model, the data. Because in the end, the AI you deploy isn't a reflection of the model you picked. It's a reflection of the data you handed it. Most of us have raw data sitting somewhere right now that forgotten folder, the drive nobody audits. The information that should have gone away years ago. But as long as nobody touches it, nobody notices. Then AI shows up and it touches everything. If you want a place to start, security, now part of Veeam, is built for exactly this. Securing your data so AI can be trusted. Next time on AI Unpacked, if you got hit by ransomware tomorrow, could you actually recover clean? Under Canada's new cyber rules, that answer matters in ways it never has before. I'm Rick Byrne. This has been AI Unpacked. Thank you for watching.
