Tech Titans: How Can a Chatbot Steal a Model's Secrets? with Steve Orrin, Federal CTO at Intel

(Federal CTO for Intel Corp at 00:00:00)
And it's that transition that we're seeing a lot of people struggle with of how do I actually get value out of this AI? And it comes back to this notion, you know, you can't just say, well, I'm gonna sprinkle AI pixie dust on every problem and magic AI is gonna solve it all for me. There was a great example. Now it's almost six, eight months ago where a bank was attacked, where a deepfake along with a chatbot was used to trick a financial analyst to transfer $25 million out.

(Joel Beasley at 00:00:33)
And so what's your official role?

(Federal CTO for Intel Corp at 00:00:35)
So I'm the federal CTO for Intel Corp.

(Joel Beasley at 00:00:37)
And what does that mean?

(Federal CTO for Intel Corp at 00:00:39)
So as federal CTO, my job is to represent all of Intel's technologies, capabilities, and solutions for the federal ecosystem. So that's the federal government, the system integrators, and manufacturers that service — we call them the defense industrial base — and the federal portions select Dell Federal, HP Federal, and the like. And my job is really to help the government adopt the technology, plan for what's coming next in the technology roadmap, understand the nuances of different architectures and ecosystems, and really as an adviser to the federal government to help them achieve their mission and enterprise goals, whether it be performance, security, how to take full advantage of AI. Because for the most part, the government doesn't buy direct from Intel.

(Federal CTO for Intel Corp at 00:01:27)
They buy through the channel and through the ecosystem. Planes from Lockheed. They buy servers from Dell and HP, and it's Intel ingredients inside. And so we work closely with the government and with that channel to help them get the best out of the systems they're purchasing and plan for the next gen of architectures. You know, you see the hype. Well, yes, you apply AI and it will solve all your problems. Well, no. That's not actually true. And we're starting to see some of that where there was a large uptick in acquisition of AI technologies, both the hardware, the software, the solution sets.

(Federal CTO for Intel Corp at 00:02:00)
And the next question that a lot of people are asking now after I've been playing with this for a couple years is, okay, where's the value? Where's my return on investment? How much utilization am I actually getting out of this infrastructure? And that's sort of the, you know, in business terms, the crossing the valley of death problem.

(Federal CTO for Intel Corp at 00:02:16)
You know, we have all these great exercises and demos and really cool widgets that we can use online. But how does that translate into something that saves me money, executes the mission faster, saves lives, whatever the industry is? And it's that transition that we're seeing a lot of people struggle with of how do I actually get value out of this AI? And it comes back to this notion, you know, you can't just say, well, I'm gonna sprinkle AI pixie dust on every problem and magic AI is gonna solve it all.

(Joel Beasley at 00:02:44)
You can say that.

(Federal CTO for Intel Corp at 00:02:45)
Yeah. You can say that. It doesn't work that way. Right. And it comes back to fundamentals. A, are you asking the right questions? Or are you defining the problem correctly in order to be able to apply the right set of technologies? And not every AI — sometimes you don't even need AI. Simple machine learning would solve it. And then the other question is, do I have the right data to drive that outcome?

(Federal CTO for Intel Corp at 00:03:10)
You know, you wanna do things, but when you actually dive into it, whether it be a security use case or a certain type of sensors, I wanna be able to detect this new kind of threat. I wanna be able to detect this kind of cancer. The question you have to ask is, do I have sufficient data to train a model to be able to augment what we're doing today? And oftentimes, you find you don't have either the volume or the diversity of data to get you where you need to go. You know, you could do a great little experimental lab, but that doesn't necessarily translate to something that can go scale.

(Federal CTO for Intel Corp at 00:03:43)
And so a lot of the challenges we're seeing people have are in getting it out of the experiment phase and into something real. They're just not seeing the return on their investments there. And I think that goes back to poor planning or buying into all the hype as opposed to what is practically achievable and where can we start first? So I think, you know, going back to your question, I think there's absolutely opportunity to AI in government and in broader private sector and industry. It's picking the right, you know, Goldilocks use cases to tackle that has the right data that you can answer a question.

(Federal CTO for Intel Corp at 00:04:20)
And then at the end of the day, when you're done, it provides value to the organization or to whoever you're servicing. And, you know, a ChatGPT isn't gonna solve a mission problem. It's a cool toy, and it can give you answers to questions in the form of Shakespeare. But can it actually give a warfighter information that they need in time of battle? Can it give a logistics person understanding of what their risks to their supply chain are?

(Federal CTO for Intel Corp at 00:04:47)
Those kind of questions require a very concerted look at how do I achieve that goal? What is the right data? What's the right implementation? What sensors do I need in order to be able to collect that data? All of that has to go into the planning process.

(Federal CTO for Intel Corp at 00:05:01)
And it goes back to what I was saying, you know, asking the right question or in the case of the prompt errors, what is the thing you're trying to get it to do? You know, if you just say, well, write me code that does all these different things and you don't give it specifics, you know, you're gonna get garbage out. But if you're looking for — and I've seen these really cool examples of, you know, hey, I need to do something weird with the BIOS. And if it's trained on the right data, so that's step one.

(Federal CTO for Intel Corp at 00:05:23)
You know? Let's say I've trained it on the specs of the architecture and the specs of the code for the past twenty years, then I can ask it a particular question. How do I enable this kind of feature in a downstream? Like, you have to be specific.

(Federal CTO for Intel Corp at 00:05:35)
And then you'll see the magic will happen. It will actually get you answers that work, but you have to ask the right questions. Or, you know, the example you're talking, you have to prompt it with the right set of queries that leads you to what you're actually looking for.

(Joel Beasley at 00:05:48)
Even I was getting frustrated with it. Right? Seventeen years as a software engineer. And then what I did was, I was like, alright. So I've got these two blockers.

(Joel Beasley at 00:05:58)
One, I'm actually trying to put it into practice, and I'm not getting the good results, but I'm really making a concerted effort. But two, I hear so many people talking about the positives and how they're crushing it and moving so fast. So I said, okay, Joel, there's a problem. There's a disconnect. To solve that disconnect, the way I did it, I watched YouTube videos of people actually building with it.

(Joel Beasley at 00:06:19)
And what I realized is I was using the tool wrong. There's like a very specific way to use it that will yield incredible productivity gains and make things super fast. But if you try to just describe a full application and hit go, it's gonna take you more time debugging it than it's worth. But if you know how to apply it and where to apply in your workflows, well, now it can take 90% of my job, and I know how to use it. So I think that we're gonna see a lot more of people getting exposed to it because we know the technology exists.

(Joel Beasley at 00:06:55)
We see highlight reels and videos and demos, but you have to actually sit down and watch someone who knows how to use it well and then copy them.

(Federal CTO for Intel Corp at 00:07:02)
Absolutely. And I think, you know, one of the things you've mentioned, YouTube is a great resource for everything, whether it's how to use a language model to write code or how do I fix the refrigerator that's on the fritz. I mean, it's amazing. But it goes back to — and this is one of the things that didn't get as much fanfare as all the other executive orders about AI was targeted funding for education and helping to train the next generation workforce on how best to use those tools. Because, you know, what you want is a workforce coming in knowing what you've learned in YouTube.

(Federal CTO for Intel Corp at 00:07:38)
And being able to then use that as the baseline. And I think what we'll see over the next several years is more, you know, earlier in people's education, they're getting exposed to the technology. They're learning it like we learned fundamental programming back in the past.

(Federal CTO for Intel Corp at 00:07:53)
And then when they get into the workforce, they're that much further along and are already, you know, the AI native, if you will. And I think having the government help drive funding for both university and K to 12 that adds AI to the mix will give us better outcomes downstream. But in the meantime, like you said, you know, whether it be YouTube videos or online courses or I find that sometimes just sitting down and playing with it and getting familiar with the tools. I like to get your hands dirty.

(Federal CTO for Intel Corp at 00:08:22)
Try it out. But, also, like you said, don't say, build me an application that does transaction processing. Rather, pick the function that you're trying to do and say, I'm trying to build a Python tool that does — you know, get very specific, and you'll find that it has been trained on so many examples that it can give you a fairly decent — now AI code is gonna be buggy just like human code. So you're still gonna do some testing there, and you're gonna wanna secure it. But it can speed the process.

(Federal CTO for Intel Corp at 00:08:51)
And I think I would offer one change to your term that you used a moment ago. It's not gonna replace 90% of what you do. It's gonna make what you do that much faster and more efficient. Understood what they could do with the technology in a novel way to go solve a problem they have. And the forestry example, and they did an experiment.

(Federal CTO for Intel Corp at 00:09:09)
It's almost five years ago now, where today, if I wanna have — if the National Forest wants to understand what the current status of blight is in the national forest, they have to send a forest ranger out on the trails, and they have to take pictures and look and see how many trees are diseased. And so you're getting a ridiculously small sample size, what you can see from the trail. And, you know, usually one or two people doing that survey, you're not getting good data. And so they looked at an approach where they use drones with cameras. They implemented an AI that were able to detect blight on trees in differing lights, different aspect ratio.

(Federal CTO for Intel Corp at 00:09:47)
They created a really good AI, worked with some universities and some commercial entities, loaded that up onto the system, and then sent them out semi-autonomous. So you've got a region, and then they flowed autonomously to be able to navigate through the environment. And they were able to get a much larger percentage, a significantly larger percentage of the forest covered. And the numbers — I don't remember the exact numbers, but the time to detection before it became a catastrophic event dropped significantly. So they were able to deal with the problem much quicker.

(Federal CTO for Intel Corp at 00:10:17)
And so this wasn't some senior leader or forester. It was someone who was looking at the problem of disease control within the national forest and had an idea of how to actually apply these technologies in a real world scenario and was able to get amazing results. And it's, you know, and then move it on from a great experiment in a couple of forests to a more programmatic approach across the national forest. Again, detect blight quicker. We're seeing that kind of thing being used now in California for forest fires and being able to detect areas before the forest fire picks up, being able to detect the humidity levels, the dryness of the leaves, the growth of the brush using camera sensors, heat sensors, and the like.

(Federal CTO for Intel Corp at 00:10:58)
So there's a lot there. There was a whole track at Defcon this past year just on AI and AI security and AI hacking, and the year before as well. Things like prompt injection and model poisoning and all those kind of targeted attacks on skewing the AI, disrupting the AI, changing what it's gonna recognize, to stealing model weights or being able to infer IP from what has been trained by using prompt injections in order to figure out what somebody may have trained it on if the data may have been sensitive or corporate IP. So there's a lot of those kind of threats that are just a new category of threat that we have to deal with. I think the other is looking at the adversaries themselves are leveraging these very same technologies, whether it be through deepfakes, really well crafted phishing campaigns, much more efficient discovery of support services and weaknesses, doing vulnerability scanning with an AI backend.

(Federal CTO for Intel Corp at 00:11:59)
And I think one of the challenges that cyber defenders have to deal with is it may take us years to really adopt AI because we want it to be secure. We want it to be trustworthy. We wanna be able to play. We gotta get budget. The adversary sees a new widget in their arsenal.

(Federal CTO for Intel Corp at 00:12:13)
They adopt. They're already out there using these advanced techniques long before the defenders are able to even get a line item on the budget for next year's budget to adopt some technology. So that cat and mouse, it's really enabling the adversaries. So they're not only targeting AI systems themselves, and we've seen a lot of examples of AIs being skewed or attacked. But, also, they're using those to prosecute their ransomware, their data breaches, and their phishing campaigns.

(Federal CTO for Intel Corp at 00:12:40)
I mean, there was a great example. Now it's almost six, eight months ago where a bank was attacked where a deepfake along with a chatbot was used to trick a financial analyst to transfer $25 million out. And it was a deepfake of the CFO with the chatbot talking to that analyst. And so they're using — that's a very exquisite attack. They had to train that deepfake generative AI on pictures of the CFO.

(Federal CTO for Intel Corp at 00:13:05)
Of course, CFOs go on CNN and talk to the press, so there's lots of media about them. And so we've seen the example of these kind of technologies being used against us. And so how do we come up with mechanisms that defend is gonna be a key question.

(Joel Beasley at 00:13:20)
The other is that — Did you say $25 billion?

(Federal CTO for Intel Corp at 00:13:24)
Million. Million. Million.

(Joel Beasley at 00:13:25)
Okay.

(Federal CTO for Intel Corp at 00:13:26)
Yeah. $25 million. That's a large number. And so being able to detect the blood flow in my veins and whether in a generated image, the blood flow isn't gonna follow the natural heartbeat rhythm. And so being able to detect a deepfake by the biomarker, in this case, of blood flow in my veins, is one mechanism that's highly accurate for facial deepfakes.

(Federal CTO for Intel Corp at 00:13:51)
Another one is the eye gaze. A lot of deepfake videos. You know, your eyes will do one of these jobs, or one will be pointing this way, one will be pointing based on what it was trained on. And so there's a lot of those biomarkers as one area of research.

(Federal CTO for Intel Corp at 00:14:04)
Another is there are certain signatures that can be detected on the video based on the generator. So, Joel, I look back at my career, and I've been doing this now for just over thirty years, not federal CTO, but being CTO of various organizations. And I think it comes back to a couple of key things. Number one is surrounding yourself with people way smarter than you and listening to them. I go back to a book and also I got to meet the author, Jim Collins, in Good to Great.

(Federal CTO for Intel Corp at 00:14:32) And one of the things he talks about, get the right people on the bus. And when he gives his talks, he talks about surrounding yourself with the really smart people in each of their domains. And I took that a step further, and you want to have a diverse group of people surrounding you. So people who are experts in a variety of technologies and business functions. And listen to them and get them to interact with each other so that you can get the best answer.

(Federal CTO for Intel Corp at 00:14:54) And so I find that having teams of people that are smart, that are experts in their domains, and bringing them together with different backgrounds and really using them as a team to solve those problems has been something that's led to a lot of my successes over my career and building up your teams and growing your teams so they become the next CTOs of whatever organization they go to next. So that's one. I think the other is, and especially when you're whether a product CTO or a customer-facing CTO, talk to customers. Get out there and speak to the people who are actually going to deal with the pain of your technology or of your ideas. Understand their needs.

(Federal CTO for Intel Corp at 00:15:32) Understand their requirements. Oftentimes, the CTOs, we fall into two traps. Either, you know, we know everything, so we're going to figure it out ourselves, or, well, the customer doesn't know what they want. They're going to want this. And there is a certain amount of that.

(Federal CTO for Intel Corp at 00:15:43) You, the CTO, you're seeing a lot of technologies that an individual customer may not see. But the benefit of talking to lots of customers across different verticals and different areas is you get to see that amalgam. You get to see what different people are struggling with and can see the trends that individually they don't. The other is oftentimes you can see where they're struggling today and figure out, you know what? They're actually going to have a bigger issue in the future, so I can help them today, but I also want to plan for what they need next.

(Federal CTO for Intel Corp at 00:16:10) And so I'd say that, you know, definitely getting out there and speaking to your constituents, speaking to customers, speaking to partners, and hearing what they're trying to accomplish, understand their use case. I think the thing that served me in my current role for the past ten years is not only understanding the technology, but understanding what the customer is trying to do. You know, the Navy has a problem, has lots of problems. They're trying to solve those problems. What is the environment they're trying to solve it in? What are the things they're trying, what's the outcome they want? And then apply the technology to it as opposed to the other way around, trying to come up with cool technology and force it on the customer base.

(Joel Beasley at 00:16:48) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.