Hello, and welcome to another Technology and National Security podcast from the Center for a New American Security. My name is Cara Frederick, and I'm your host for today. We're here to talk about the launch of our new AI report series, which we began with our first report, Artificial Intelligence, What Every Policymaker Needs to Know. You can find this report along with all of our AI content at www.cnas.org.
And we have with us today Paul Shari, Senior Fellow and Director for the Technology Program here at CNAS, Greg Allen, Adjunct Fellow on the Tech Team here at CNAS, and Eduardo Saavale, researcher for the Energy, Economics, and Security Team at the Center for a New American Security. Paul, we'll start with you. Can you tell us a little bit about the reports and how they fit into our broad AI and Global Security Initiative? Yeah, we made a decision last year to launch this new artificial intelligence initiative and to focus the bulk of our research here at the Tech Program on AI because it is such a transformative technology, it's seen such really incredible gains in the past couple of years, and it's clear that it's transforming a range of industries.
And I wanted to really help spark a disconversation in the national security space about how this will affect AI. I'm very excited about this report series that we're releasing this summer. Digging into this issue, you really talk about how will artificial intelligence affect international security and table to release these and work with Eduardo, Greg, and many of the other co-authors on this really interesting series. So let's take it back to basics.
Can you tell us what artificial intelligence is? In its essence, AI is a field of study thinking about how to make machines that do intelligent things. Now, what is intelligence? Is it debated concept?
There's a variety of definitions. We certainly define the term and we pick one that works well for policymakers in our first report in the series, whatever our official intelligence, whatever policymakers needs to know. But it's basically a set of cognitive processing tools that allow machines to do interesting and clever things, process large amounts of data, solve problems for people, and perform tasks that otherwise humans would normally do. All right, and last basic question before we dive into the substance of our second report, which we are launching soon.
So what is machine learning and how does it fit into artificial intelligence? Machine learning is a method of creating intelligent systems. Not all intelligent machines use machine learning, but many of them do. It's particularly a method that's really taken off in the last several years because of the ability to have large data sets that can be used to train machines and increases in processing power and better hardware that can then process this data and use it to train algorithms.
And so we found that many problems where it might be hard to write a set of rules to tell a machine what to do. For example, distinguishing between visually and Apple and a tomato, both round and red and shiny and blue-green stem on top. You can train a machine to tell the difference the same way a toddler can, simply by looking at large volumes of images of these things, a machine can learn to cluster them into different groups and tell the differences in them. So what I'm primarily interested in, and what I think a lot of our listeners here in the national security community might be interested in, is the second report in the series.
So that's titled Artificial Intelligence and International Security. And Greg, I have a question for you given that you wrote so extensively about cybersecurity and its nexus. Can you talk a little bit about the intersection of AI and cybersecurity for us? Sure.
In October of 2016, I was at an attendance at a speech given by the then head of the NSA Admiral Michael Rogers. He said that AI and machine learning is foundational to the future of cybersecurity. He said that only a few months after the DARPA grant challenge on cyber was using AI systems that could automatically hack into each other and try and disable each other and at the same time detect vulnerabilities in their own systems automatically and try and patch those vulnerabilities. That level of automation was to a much greater extent than we are familiar with seeing in the cybersecurity landscape.
Though cyber tends to be less labor intensive than other forms of espionage and attack, it is still very labor intensive and especially very high skill labor intensive. And what we are now seeing with advanced artificial intelligence technologies is that certain categories of labor automation and cybersecurity are now available in a way that they were not previously. So this is an exciting time. It's an early stage that original DARPA competition used a lot of automation but not a lot of advanced machine learning.
I think if you were to run that competition again today, you would see much more extensive use of machine learning in order to automatically detect certain types of vulnerabilities and to exploit those vulnerabilities through attack. Can you talk a little bit more about the applications of machine learning in the cybersecurity domain, especially as they could potentially affect international security? I think there's a few different buckets of the implications I would say. The first of which is the advances in artificial intelligence are increasing the connectivity of physical infrastructure to digital systems.
This is a trend that predates the rise of artificial intelligence but its implications are much greater. So the more activities that you are using computer systems to do, the greater the consequences of hacking it. So if you are, for instance, hacking a traditional automobile, you might be able to turn off the brakes. But if you were hacking a driverless car, you could drive it off a cliff.
And so the growth of artificial intelligence increases the potential consequences of traditional cyber hacking. The second is that artificial intelligence systems using machine learning have a different programming architecture and therefore different types of vulnerabilities than traditional software vulnerabilities are hacking. And so the mechanisms of attack might be different. The few of the ones that we talk about in the report by the idea of data poisoning.
Since in a certain sense, the software is programming itself based on exposure to training data. If you have access to that training data, you can insert examples that will lead the system to make predictable types of mistakes. So these are, this is just one example of new types of cybersecurity vulnerabilities that are opened up by greater use of artificial intelligence. And then finally is the opportunity to create cyber offensive and defensive tools using artificial intelligence technology.
Again, I would say that we are in the early stages of this, but some early, some early, some early results in the research community that are quite interesting is the use of automated fuzzing in order to automatically detect vulnerabilities. So normally when you are going through a software piece of software and trying to detect cyber vulnerabilities, this is a pretty labor intensive process that requires real expertise and a lot of time. Now, because certain types of vulnerabilities are frequently identified, there's now an opportunity to automate the process of this vulnerability detection, which is really quite useful if you are trying to deploy software and you want to have some confidence that is not going to have the same types of cyber vulnerabilities that we have identified for decades. That's actually one of the more interesting features of cybersecurity is usually it's the same types of vulnerabilities that have been known about for a long time that just keep popping up integer overflows or other types of vulnerabilities.
So cybersecurity offers that defensive opportunity. But then in the offensive realm, there's the opportunity to run machine learning algorithms on previously successful forms of malware, to make slight modifications to them, to run simulations to test those and to do so at scale. So it's great viruses that we'll be able to get around to new security patches at a much greater speed. So all those are interesting and you've alluded to the fact that some of them have been done and they're just cropping up again.
But what are we not thinking about? What's new with regard to cybersecurity and AI? Well, as with everything, I think what's quite new is the applications of machine learning. So in traditional software programming, there is a person who is typing in every line of code.
And every application of the code is some sort of long sequence of if-then statements. If you are exposed to this input, then provide this output. With machine learning and it's a bit of an oversimplification, the software is to a certain extent programming itself by applying an algorithm to a training dataset. That story is what allows for the radically improved performance in automation and the automation of certain categories of tasks that could not be previously automated.
And bringing that into the cybersecurity domain, which is an incredibly data-rich environment, offers some real new frontiers of automation throughout visual intelligence. And Eduardo, despite my inability to pronounce your last name, you talk extensively about economic and financial tools of statecraft. And in doing so, you assert that AI tools can potentially strengthen counter-listet finance operations. Can you talk a little bit more about that?
Yeah, certainly. So since 9-11, the United States has created this major global apparatus to counter illicit financing. This means anti-money laundering rules rose to counter terrorist financing and sanctions. These have been very useful tools.
However, even to this day, they are very leaky. A UN study estimates really catch 1% of illicit finance flows coursing through our system. What I can do is it can really strengthen our systems in a variety of ways. The first level, just the pattern recognition level, on the pattern recognition level, we can create much finer tools to find these illicit financing flows.
So for example, when we deal with North Korea sanctions, banks are to, quote, watch out for transactions that involve China. That's a huge filter that is very useless from actually finding these flows because it generates an absurd amount of false positives. By using unstructured datasets, we can find new ways of slicing the data that human and the compliance officer couldn't use to find these patterns. For example, in one case, looking through bank transaction data, we found that using the browser language was a very good predictor for fraud.
So illicit actors were changing the browser language to try the full compliance systems and then the AI system found a strong correlation with fraud activity. On the second level, we can incorporate completely new data that a current system, especially a human centered one couldn't use. So for example, there's huge amounts of data for open source data, say on social media, that current compliance operations don't use. An AI system could train this and combine it and bring new sources of data that would lead to far more robust compliance programs.
Additionally, using an AI could, an automation could create a better cross-spectrum analysis of threats. Currently banks are called between a rock and a hard place. On the one hand, they have to fulfill all their obligations to the US government. So check for all sorts of threats, whether it means blood diamonds and trafficking in Africa, but also the wrong sanctions.
However, so often they adopt kind of a check the box approach where they do the minimum they can to avoid legal responsibility. At the same time, they're also playing catch up with the government to respond to its shifting priorities. So for example, on May 8, the Trump administration announced it would be re-imposing sanctions on Iran. So all of a sudden compliance programs, even though they technically kept looking, will have to reorient themselves to make sure they're not exposed to major legal liabilities as for example, just as the department looks to levy fines and banks that don't strongly enforce Iran sanctions.
What AI can do is, like increasing the automation, as part of this, it can always keep an eye on these things and get rid of a seesaw effect where banks are trying to guess what regulators will want from them and instead can ensure a more level set approach to finding legal transactions. And I guess the final thing I would say is that there can be a huge reduction in costs. Ever since 9-11, banks have been given all this responsibility by the United States government. They are the front line of say fighting terrorism or enforcing US nonproliferation policy.
Large banks, in part, motivated by major fines because they initially didn't do it, have been able to stomach these prices. Smaller banks have not because they just either don't have the know-how or don't have the resources in place. This has created very weak access points in the financial system, which you see are constantly exploited by terrorists, by North Korea. By lowering the cost of these compliance programs, AI can ensure a stronger system all across.
And the number one way they can do this is by lowering false positives, which are the main thing that results in major use of manpower hours and deflects from a more effective AI, effective compliance system. So that's a pretty full-sum enumeration of how AI can engender innovation. But what can some tools do to disrupt the counter-elicit financing framework itself? I think the main thing when we talk about disrupting the counter-elicit financing framework is it could really change the way we approach it.
Although there are general indications, the main way we do counter-elicit financing is through fairly rigid rules. So on the most basic level, the UN and the United States published lists of sanctioned individuals and banks can't do business with these. This creates clear-to-date catch-up environments. The United States is always trying to list new people and then banks need to update the list within their filters.
And you're always a step behind. Even more generally, there's a set of very hard and fast rules. So for example, a transaction that involves more than $10,000 triggers a report. A smart person can get around these systems very easily, especially once their sophistication increases.
So what major international institutions have been calling for is a so-called risk-based system rather than a rules-based system. So we're moving away from rigid rules to a constantly updating mitigation and adaptation system. It is very difficult. And especially as I was mentioning earlier, banks might feel they can use a risk-based system in one against one threat.
So say, counter-elicit financing, but in another threat, they might not say nonproliferation. What we can do with AI is directly move completely to a risk-based system that will completely reshape the way we think about counter-elicit financing. And then, Paul, I'll move back to you as the mastermind of this global security and artificial intelligence initiative. So I have a broad question about national security and this community itself.
How do we prepare for the consequences of the AI revolution? Thanks. Yeah. And that's my two-speed overblown.
But I'm glad that we were able to build a great team here at CNAS, people who were thinking about this. The way that we've been talking about it, that lots of people are talking about AI is something that's on the scale of another industrial revolution. I think that makes sense when you think about the scale of potential change that the technology can break in terms of changing economics, changing industries, we're shaping politics, society, international relations. And we've seen from prior industrial revolutions that they changed the balance of power among nations.
Nations rise and fall, and even change the fundamental building blocks of power. They changed the key metrics that matter during the industrial revolution. We moved from an era where arable land and population were the key determinants of national power to a place where coal and steel production and oil production all became major determinants as well. And it shifted to geopolitics.
Oil became a global strategic resource. And so now we're in an era where new tools of power matter, things like data access to good quality data, protecting that data, data integrity. Hardware is absolutely critical. And in particular, you know, specialized hardware that might be used for machine learning.
And so now we're in a new era where nations are competing for data, for hardware, and for human capital, for talent, and the people that are going to be designing these algorithms and employing them. And one of the reasons why we're releasing this series, because it's really important that the US remains at the forefront of this global competition. So I want to probe one of the things that you talked about there. And that was the fact that AI will be changing industries.
So what do you see as an impact of AI on the future of work? Are we going to see Luddites revenge? Can you explain what that even is? Yeah, we walk through in reports of a variety of possibilities.
There are just a wide range of possibilities and predictions about what AI will mean for the future of work. Anywhere from, you know, there'll be massive layoffs and something on the scale of a great depression. And people learn jobs too, on the exact opposite people saying, well, I will create all of these new jobs. You know, when you look at them, some of the most convincing deep analysis, I think comes from an McKinsey group that has basically suggested that somewhere about 45% of jobs in the US today, or job tasks rather.
So job tasks being done in the US today, or have the potential to be automated. But it's only about 5% of jobs that could be completely eliminated. So most jobs will be transformed as many routine physical and cognitive tasks that people do will be offloaded to machines. Very few jobs are eliminated entirely.
Now industries may consolidate positions, companies may consolidate shrink the workforce, but new jobs will be created. The challenge, of course, is are those the same types of jobs? Do we have people trained at the right types of skill sets? One of the things that we did that was new, that I've not seen elsewhere, is compare that analysis by job task with the Bureau of Labor Statistics analysis of the US workforce in different industries by age.
And we found that the automation is likely to hit younger workers the hardest. Now on one level, that's good news, in the sense that they have the most possible time to adapt to a new changing economy. I'm certainly, for somebody who's in their 50s and 60s, it's much harder if they're later off for them to be able to try to retrain and acquire new skills and be competitive in a workforce. The flip side of that is, you know, for many folks, if what's really needed to stay competitive is longer education and more education, it's going to take more training for them to really get to a place where they're performing higher level cognitive tasks.
If we're taking away some of the routine labor jobs, the barista, things that are allowing people to pay their way through college, that could begin to widen income inequality and make it hard taking away some opportunities for people. So we want to think about, as we move forward in the coming decades, public policy choices, to make sure that we're giving people the right tools and education to stay competitive and what's going to be a very dynamic and changing workplace. So we'll wrap up here in a second, but my last question, and this is for all of you, what is the most important aspect of artificial intelligence and international security and the way that relate, that they relate, that we're not quite thinking about right now? Paul, we'll start with you.
You know, I think one of the most interesting pieces of this technology that is really critical for policymakers and national security space to understand is that a machine learning base tools work very different than many of the rules based systems that we become very accustomed to, you would understand, they're going to be tuition about as we work with computers and software and as we're over day to day lives. They have different strengths, different vulnerabilities, different weaknesses, and in particular, the national security space, we have adversaries that might be trying to hack them to poison a data, to manipulate them, to trick them, we want to be cognizant of them and make sure that we're robustly testing our systems ahead of time because we're in a very competitive space and we want to rent in and know what our adversaries are going to do before the government. And why don't we share a plug? I think I would just say that the United States, currently as the central global, central to global financial system, has enormous power and this power allows us to impose rules and collect huge amounts of data and the United States policymakers have gotten used to collecting all this data and forcing people to disclose it, but have really not used it as much as they could today.
So I think one of the most interesting things we can expect from AI is all of a sudden realizing how much data we have that we didn't even know we had or we didn't know we could use and starting to employ it for something that already has been successful to this day. So there could be the whole new horizon of success within the financial integrity space. And Greg, you have the last word. I would say the most interesting aspect of artificial intelligence is what it is doing to reduce the barriers to entry to certain levels of technological capability of international security relevance.
The example I love to use is that of the Tomahawk cruise missile. This is a system that is capable of delivering an explosive payload over a range of 700 or several hundred miles to an accuracy of several meters. That is, but the downside is that it costs $1.5 million per shot and that you need thousands of very high quality aerospace engineers in order to design that system. Now the United States hates paying for this system, but it loves that the system is expensive and it loves that the system is incredibly technologically complicated because that plays to the United States' strengths in the international system of having a lot of money and having a very high quality educated workforce upon which to develop these systems.
I think the challenge of AI is that it's going to lower the barriers to entry for these types of capabilities. If you think about autonomous drones that are coming off the commercial market, they are offering not Tomahawk-level capabilities but something vaguely approximating that at a tiny, tiny fraction of the cost and with commercial off-the-shelf technology. So you do not need dozens or hundreds or thousands of aerospace engineers to have these capabilities. And I think leveling the playing field and getting certain types of actors access to capabilities that they wouldn't normally have in the historical sense is really quite interesting.
If you think about the fact that ISIS, an insurgency was able to use drones for overhead reconnaissance for dropping explosives. We don't normally think of insurgencies as the type of groups that have air forces. But that is what commercial AI technology is enabling, is more actors to have more capabilities that historically would be enjoyed by large, powerful states. And that's a trend that we're going to be wrestling with for decades to come.
Alright, thanks all of you guys. That's all from us here at the Center for a New American Security. And once again, you can find all of our AI content at www.cnas.org. Thanks again.