Wild Dog AI Podcast | Ep 7 | The Race Against Adversaries: Embracing Tech for Enhanced Securities
Welcome back to the Wild Dog AI podcast. In our last episode, we spoke to Randy Stickley and we talked about the need for human ingenuity and things like critical thinking to maximize our use of language models and generative AI in the law enforcement and intelligence space. Today, we're going to jump into the adversarial and bad actors use of language models in today's world and what it might look like in the future as we move forward. But before we get started, we have Heather with us today, who you've met on previous podcasts. And Heather, I just wanted to sort of talk very briefly about how language models are manipulated, because it's not the same way that traditional software applications are manipulated with things like malware. Adversaries and bad actors are constantly coming up with new ways to leverage this technology to manipulate.
Not just our software applications, but also society through scaling propaganda. And we'll dive into that a little bit too. But for the audience, just to kind of put it into context, the way I think about the way I think about the manipulation of language models is it's almost like we're dealing with a child right now. We're dealing with a five or six year old kid that has some parameters. They know what's right. They know what's wrong. But with language models, they're easier to manipulate because
They're constantly getting new instructions. They're constantly training. They're constantly being updated, just like a child is, right? You can tell your five -year -old kid what's right and what's wrong, but then they go out and they're exposed and hang out with their friends and they learn from other sources and they sort of, those parameters change. Those guardrails widen just a little bit. They widen a little bit more and over time, you know, it's the parent's responsibility to go in and kind of reduce those guardrails and get the kid back on track and explain, no, that's actually wrong. That's not good.
And so some of the recent attacks that I've been reading about using language models is for instance, people uploading PDF documents into a language model. And in that PDF document, there may be instructions for that language model to, hey, this is, these are some new instructions on how you should operate, right? Or, hey, I'm a good guy. These are some new instructions Buddy (10:42.124) your output or how you
should reason or the perspective you should take when you answer these questions. And so I think really what that means is we've got to start thinking about when we introduce new data to our special purpose models, if we're building our own language models, we've got to be careful of what data we're uploading and introducing to that language model. Just like we've got to be careful what we're exposing our five year old kid to, right? Because they're constantly learning, they're constantly evolving, they're constantly those guardrails are constantly shifting. And so I think that's one of the ways that language models are currently being manipulated, especially when we talk about propaganda and scaling propaganda surgically. That's one of the ways I've seen some recent attacks. And so the manipulation of the APIs and how those APIs are configured, I think is extremely important.
For instance, if you configure your special purpose model and the API that you're connecting to it has both read and write privileges, the chances of an attack or the chances of your model being manipulated are much higher than if, say for instance, it only has read privileges as opposed to write privileges. And so there's a of things there that can actually be done to reduce the amount of threats or increase the security of these language models. But it is kind of a different way of thinking about cybersecurity. I think it's a completely different way of thinking about how we train these models and how we ensure that we keep them safe for our end users. And then obviously, you know, it's inevitable that we do have to develop a little bit more skepticism when dealing with media today, because to me, it's inevitable that social engineering is only going to get It's not going to get worse, right? We're developing deep fakes, creating propaganda. Those things are only going to get better. They're not going to get worse. But I think there's a whole new industry now for how do we combat
those things? How do we truly understand what the difference is between an AI model building something and truth. So I think there's a whole area for that. So Heather, I just wanted to sort of set the stage to kind of give people a way to think about Buddy (13:04.098) in terms of LLMs and how they are manipulated versus the, you know, traditionally how applications were manipulated with malware and so forth. And the way you manipulate a language model is
a little bit different. And the approach is a little bit different. How we think about it is a little bit different. So let's jump into some of the more recent examples of how adversaries are using language models and manipulating language models to conduct attacks. Heather Perez (13:33.334) Well, just in general, like some of what you were mentioning is referred to as like data poisoning to get the outputs that you want from the language models. And a lot of people when they're thinking about the AI and generative AI, they're thinking about ChatGPT and more recently all the improvements in everything that's come out over probably like the last year and a half or so.
But there's been AI related bots and all of that for a long time. There was an incident back in 2016 when X was still called Twitter, where Microsoft had launched a bot called TAI. And it was meant to engage with millennials, like from teenage up until like in their early 20s. And within a short period of time of that bot being launched online, it had been manipulated in the responses that it started giving was like incorrect political information, racist comments, and it just kind of went rogue. It just went off and started talking about everything that it was not trained to do based on how other people were engaging with it. And that's kind of the method when you're looking at some of the, especially some of like terrorist groups and extremist organizations, cyber criminals, that's kind of what they're doing with it. They're jailbreaking it and kind of to learn, learning how to get the information that they want
out of it. So you have them, a lot of the things that are being utilized right now are deep fakes in particular, deep fakes for disinformation, deep fakes are being used for fraud schemes. propaganda, you have cyber criminals using it to amplify their malware, their social engineering, their phishing attacks, you've got AI powered malware, you have cyber criminals that are making their own chat bots that are, you know, it's like going to ChatGPT, but they've taken off all the restrictions, so it's free reign and they can do whatever they want with it. So they've made versions called like evil GPT, worm GPT, fraud Heather Perez (15:40.662) And you can look at some of those. Like one of the more popular is Worm GPT. If you go to that one,
you can see that it's for popularity, it's got 8 .5 million that have used it in conversations, almost a million conversations. So they're using these quite frequently and they're updating them. They're advertising them. One of them, Fraud GPT is advertised as your cyber
criminal co -pilot. So they're tweaking these systems to work, to work for them. Buddy (15:52.147) wow. Buddy (16:12.73) Heather, you talk about just so everybody understands I think the term jailbreaking what does that mean in terms of language models like how do you jailbreak a language Heather Perez (16:26.947) I would so for how the cyber criminals are modifying theirs I can't say exactly but for some of the uses so if I go to ChatGPT and I'm asking it to provide me information on Heather Perez (16:43.156) I'll use, trying to think of a good
example. Okay, so I'm going to ChatGPT and I want to know, I'm an extremist or I'm just a regular person. I want to just talk about insurgencies and just talk about general stuff like It's not going to want to provide the information, but if you modify your response where you're asking as a researcher or I'm trying to learn indicators for because I'm a law enforcement officer and I need to understand how to develop countermeasures. If
you tweak your verbiage enough, you can make it give responses that it doesn't want to provide. Buddy (17:16.656) so back to our sort of analogy with the child, right? Like if I tell my five -year -old, don't talk to strangers, well, they're not gonna talk to strangers, but if that stranger shows up with something that's recognizable to them or ice cream or candy or something that they want, they're gonna assume that that person's friendly. That person's no longer a stranger because they said, hey, your mom told me to meet you here. Your dad told me to pick you up at four o 'clock because he's stuck at work. Heather Perez (17:46.208) So then it
Buddy (17:46.343) So it's kind of the same thing, Heather Perez (17:50.068) Yes. Yes. So that's what Buddy (17:52.78) Yeah, I heard an example. Heather Perez (17:57.386) Go ahead.
Buddy (17:59.148) Yeah, I heard an example of like, there was an example I was listening to where somebody logged on and asked GPT -4 to help them rob a bank. And it said they couldn't help with that. And then it said, hey, I'm a law enforcement officer and I'm planning, I'm basically building a threat vulnerability assessment for a bank. So help me understand how somebody might rob this bank and gave it some details. And then it laid out a whole plan on a bad person, a bad actor might rob the bank.
Heather Perez (18:32.726) Yes. So that is one of the main ways that they are exploiting that. with ChatGPT, particularly the beginning of last year and maybe the end of 2022, like amongst the first people that I saw playing with ChatGPT and how to get around it were the terrorist organizations and racially motivated violent extremists getting on there and playing with it. Even in right after the January 6th, You had a lot of them that were going on because you had like the DALL-E bot for making images. They were going on there and using prompts to make images of like the Capitol on fire or politicians being hung and using that as propaganda. Buddy (19:19.798) So that's an interesting
part of how they're leveraging these language models too is so it's not to just generate the propaganda or generate the images or the text. Can you talk a little bit about how they're using it to propagate that messaging in those texts and how, how uses, how the use of things like language models can help them actually proliferate those messages as Heather Perez (19:45.098) Well, over the last couple of weeks, think it was around, it was July 9th, there was a statement, press release from the Department of Justice about a state -sponsored Russian bot farm that was disrupted. So the Russians have had bot farms on Twitter and now X.
And that was kind of a normal thing. for this particular one, they found that they were doing using AI enhanced bot farms that was allowing them to run numerous fictitious accounts. And they were creating authentic appearing social media and personas in mass. That's kind of one of the indicators. If you see like a network of accounts that are spreading the same messaging, they're deploying content that was similar to what you would see everybody else using.
So they'll they'll use similar propaganda and maybe tweak it a little bit. If you look at the networks, like if you start looking at one account and start seeing the same thing, you could usually track those accounts by who's following each other. is at least in the past, could identify the bots because they would all follow each other. They were mirroring disinformation and other bot personas. They were using false
narratives. They were propagating false narratives, formulating. Buddy (21:10.126) Oh, that's really interesting. That's interesting because I know that when we did research on the internet research agency before, one of the things they were good at is, I think there was a point in time where there were just over 1,000 employees at the internet research agency. And we discovered that through research, through publicly available data, that of those 1,000 plus employees, Each employee maintained anywhere from eight to 10 US based social media accounts. So one employee could be like, you know, Bob from New York that who, you know, leans left. could be, and they could also be Mike from Los Angeles who leans right.
So they managed on a day -to -day basis, all of these different personas and each of those personas were used to either amplify events happening in the United States. and make sure that those events were going to the people that they would antagonize or to the people that would like get upset about it. Or they were basically creating their own events. And so I think with what you're saying is kind of interesting because what
we could discern through those accounts is because of human nature, you could see that, okay, this account probably belongs to this person because there's, because as humans, we do have patterns in how we build the accounts. have patterns You what we say in these accounts, you have patterns on like the timing of how the messages are released. And so I'd imagine with, with, the ability of using AI, you could really get much better at ensuring that those accounts do have their unique identities that you could converse with a language model to ensure that you're maximizing the use of that account to achieve some sort of goal or objective. So I guess that having a co -pilot. kind of like a propaganda co -pilot could make their operations much more effective. I mean, they were effective originally,
but I guess now that you use the way you kind of laid out makes it that much more dangerous. Heather Perez (23:17.376) Well, in this particular instance, they were using an AI enabled bot farm generation and management software to help them spread the disinformation, manage the accounts. At the time, they only found it operating on X, but they said that the analysis
showed that the software's functionality could likely be expanded to other social media networks. Buddy (23:45.57) Wow. Heather Perez (23:46.26) And so again, they were already good at building these bot farms and putting out disinformation. I know in the past,
in my previous job, when I was looking at extremist groups and such, there were even instances where the Russians had created fake Facebook pages or fake accounts on Twitter where they were propagating narratives that went into kind of like... martial law and the government taking over where they actually scheduled an event, a fake event that people showed up to. So being able to now leverage a language model, they're able to put out much more targeted information and to do it a lot quicker. And when you're flooded with all of this fake information, it gets kind of hard to kind of sift through that. And also,
I mean, right now with just the internet and social media in general, it's, when you have disinformation and misinformation, it's very difficult because I may go in saying I wanna learn about something, but I actually just wanna learn. I just wanna reinforce what I already believe. So there's ample amounts of information out there now that can just kind of build and make me feel like my false narrative that I believe is actually real. Buddy (25:02.294) Yeah. And you can do that at scale now. You can sort of do that at scale. So as, as law enforcement intelligence professionals,
how do we even start to, how do we even start to combat this? How do we even start to think about, you know, cause one of the things I struggle with is I think that the, the way to combat misinformation is you've got to get the truth out there faster. Heather Perez (25:05.941) Yes. Buddy (25:31.404) Right? We see it with events all the time where a major event happens in the United States and, you know, the, law enforcement agency responsible for investigating the case and, rightfully so they, they take a little bit of time to get the truth out there. But I feel like in today's digital age, if, somebody comes out with the narrative or two narratives or three narratives before that investigation is complete, you've already captured the attention of people. You've already got people believe in what you're saying. And so sometimes.
I, you know, sometimes people don't even stick around for the results of the investigation because they've already found, found their echo chamber online, what they've read of aligned with their confirmation biases. and they've already solved the problem. So it really doesn't matter what law enforcement says anymore than that belief is sort of solidified in their brain. And an interesting thing happens when a person hears something that already aligns with their, preconceived notions is they start to take ownership of that. And then they start to defend it. And then once you start to defend. you know, that misinformation, it's yours at that point. So now, no matter what law enforcement says, is you're, you you're, going to defend what you believe to be the truth now against whatever law enforcement says. And I think that's where a lot of conspiracies start,
start to be ginned up because how could law enforcement decide that that's the outcome when clearly this is what happened, right? Cause I've spent the last three days convincing myself with like -minded people that this is what actually happened. and That to me, the one of the solutions has always been, I know it's easier said than done is how do we get the truth out there faster into people's hands and not sort of seize that terrain to these adversaries inserting false narratives into the digital domain? It's hard to do. It's really, really hard to do. So absent that, what other, what other ways can law enforcement absent being able to get the truth out there faster and conduct an investigation in record time, which obviously is not what we want to do because that's where I think that's where a conspiracy theories actually probably become more real, right? Because if we're doing haphazard investigations, just to get the truth out there, there's going to be a lot of holes in the investigation. it's, it's a, it's an interesting, it's a tough position that law enforcement finds itself in. And so you, as a law enforcement analyst, what are some ways that we even start thinking about combating bad actors ability to at scale, Buddy (27:58.434) very surgical narratives following major events.
Heather Perez (28:03.775) Well again, it's a difficult thing to do because like you just mentioned, for analysis, for investigators, intelligence professionals, you have to do some type of an investigation to start with. You're not going to have the accurate information, the true story up front. So no matter what you're going to be battling this false information. And that also makes the investigation more difficult. So while we're trying to figure out what actually happened, you're having to sift through all of this disinformation and misinformation that's being shared around everywhere to find like the nuggets of real information that are what you need to move your investigation forward. And it gets really difficult because when you have these, when you have an incident and you start having people reporting and sharing disinformation.
So let's say you have a, you know, a nation state or a terrorist organization, extremist organization that decide that they want to do a, you know, a hashtag campaign or something on Twitter to kind of, you know, put out the false narratives. At some point, depending on what it is and who's sharing it, if they can make the accounts look like they're not coming from extremist accounts, you'll have some of that disinformation. Like you mentioned, people are gonna go look for what bolsters their beliefs. So then you have what was maybe an AI enhanced
bot operation. Now the information those bots were sharing are now being shared organically. by folks that just have decided they believe that narrative. So then you're trying to track all that back and it it starts spreading so quickly that sometimes it's hard to pinpoint the information back. takes a little bit of time. And then by that time, it's like grown out of control. And Heather Perez (29:50.954) I mean, one of the main things
is that you just have to understand all of these different technologies. So the fact that nation states and terrorists and criminals are exploiting AI shouldn't be a surprise to anybody because they're always looking for new technologies to kind of move move their organizations forward. So like if you take the example of like the Islamic State, over the years since they really came out big on social media, they're always amongst the first to test new platforms that come out, new mobile apps, new online platforms for sharing information, because they want to make sure that their ability to share propaganda and recruit isn't hindered by their ability to stay on a certain platform. So, like right now, Telegram is still one of their favorites because they're still able to maintain a presence on there, but they're still always looking for other ways to bolster that. So in some of recent propaganda for some of their media groups, looking for folks that have the abilities, AI. to leverage AI is something that they're looking for because they're looking for different ways to stay online and to be able to share that information.
So you need to understand how the groups operate and how they could potentially leverage this different type of information and where they might pop up. So a big discussion amongst extremist groups is they've put a lot of effort into a lot of discussions across various different ones that I've seen about the importance of targeting the youth. So they are trying to get a presence on different platforms that you see younger people on. So you'll see a lot of them on like let's say TikTok. So you can go on to TikTok. Heather Perez (31:29.814) and the Islamic State shares videos, some of them are real battlefield videos that they label as AI, even though it's not, to try to get around the algorithms that suspend accounts. But they've also started generating a lot of propaganda that is AI -generated propaganda, just because it's quicker.
And so they're doing it just to be able to get stuff out quicker, but also to try to get around some of the algorithms that suspend their accounts by labeling it as, you know, this is AI generated, it's fake, these are actors. Buddy (32:02.367) Yeah, one of the things that's always concerned me with the way religiously motivated terrorist groups operate is, you know, the goal of a lot of their propaganda is ultimately recruiting. And I think people look at this too simply sometimes. It's like, hey, you know, this terrorist group is They're proliferating this video or this message. And we look at it, you know, an American looks at that and says, this makes no sense. This is stupid. There's no threat here. But what we
don't understand is that message wasn't intended for us, right? That message was intended for 18 to 20 year old men who live in North Africa, uh, who can't get ahead in society because they're being oppressed by the government. Um, and so that, that, that always concerned me, especially when, like in the case of ISIS, when their numbers were going through the roof and we're A lot of their messaging isn't making any sense. Well, the reality of it was, is because none of that messaging was intended for a U .S. audience. It was intended for very specific pockets of different societies, low income societies across North Africa and the Middle East. And it was very,
very effective. And I could imagine with a technology like GPT -4, where it can produce messages in almost any language very, very quickly, the ability to scale and reach more people is kind of scary. What I wanted to ask too, Heather, so it sounds like law enforcement has to continue to do investigations, right? They've got to take their time and find the truth. That's important. The rule of law is incredibly important to us. And law enforcement analysts and intelligence analysts have to still follow their processes.
I think we're doing a better job of bringing technologies into those workflows to make do it more efficiently and maybe faster. But dealing with that narrative, I know in the army we have, the army has a thing called public affairs or the public affairs offices, a lot of businesses have them. I think law enforcement has them too. Are we using our public affairs capabilities and law enforcement to their full extent? So combating those narratives in that time where we're trying to find truth.
Buddy (34:24.27) How do, what is, is that a public affairs job in law enforcement? Whose job is that? Is it the sheriff's department? Is it the sheriff's job? How do we start to think about Heather Perez (34:33.76) mean, would depend on who was running the investigation, whether it was a local agency, the Sheriff's Office, but they do utilize their public information officers. Pretty much all the agencies have social media present. So as they're able to release information on an ongoing incident, they do share information with the public. Again, it's not always gonna be as detailed as what people want.
So a lot of times what they're seeing in the disinformation is going to be more detailed than what's coming out from the agency that's doing the investigation because they can't share everything. It's not a matter of trying to hide the information, they're trying to maintain the integrity of their investigation. So it's just, it's gotten to a place now where everything is so accessible, there's so many different platforms and there's, you know, You can even go to like very niche platforms that just focus on certain topics that it's really hard. If someone wants to believe a certain narrative, it's going to be very difficult. And I can tell you as someone who was an extremism analyst,
it got very difficult because you have, especially with all the disinformation that can now be amplified by AI, have you have individuals that are not extremists that ending up in extremist networks on X or on whatever platforms they are because some of these, because of the amount of disinformation, a lot of that is what used to be conspiracy theories and kind like segregated away has mainstream. So you have mainstream false narratives. So you have people that weren't usually exposed to that now seeing that information in mass. And it's more difficult to counteract that because there's so much of it out there.
Buddy (36:30.594) Yeah. Well, I think I, you know, one of the things I do feel good about is, you know, I have, I have, I have three daughters and I think because they've grown up in this digital environment, I do notice that when, you know, they see something on one of their social media feeds, they are much more, I think attuned to separating like fake stuff from real stuff. I'll even look at some stuff and I'd be like, my God, I can't believe that happened. And know, my daughter will say that that's Well, how do know that's not real? And she'll point out why it's not real. And so I do feel good that I feel like because these younger generations are growing up on these platforms, they are developing like sort of a superpower to differentiate truth, know, facts from fiction. Not everybody can do that. You know, the older generations may
struggle with that a little bit, but I do feel like there's like this sort of inherent skill set that kids are developing because they're exposed to so much data so quickly. Nowadays, I think we talked about it before, like, what is it the average person now is exposed to more data in two days than somebody in their entire life prior to like 1975 or something like that. So there's just so much information being thrown at these kids. And so it and that's why I'd imagine that it's even more difficult to spread pop. It's going to get more difficult to spread propaganda because it's got to really be good. Whether it's truth or not,
it's got to really, really be good for somebody to willing to be beat. for somebody to be willing to share that information. So it can't just be bad information. It's gotta be attractive. It's gotta be pretty good. Heather Perez (38:06.868) Yes, but what you're saying is pretty good is also going to be like that. What's considered good.
information, it will be different for different threat groups. like if you have, again, going back to the Islamic State or, or al Qaeda, but mainly the Islamic State. So when they were putting out their official propaganda, it was always, you know, it was very high quality, very high quality, their official stuff. But when they started losing territory over in Syria and Iraq and stuff, and some of their media folks were taken out, then they began to empower their supporters. So you would, because supporter propaganda, you would look at it and be like, going to be inspired by that? That's terrible. It looks like someone just threw it. But that became when they started empowering the supporters. was more about the quality
might not have been as great as what was coming out before, but you're supporting their organization and you're putting forth an effort. You're actually doing something. Heather Perez (39:05.335) But they have a lot of discussions amongst themselves about what narratives they want to put out. So again, they're doing more targeted propaganda, targeted recruitment. And like I mentioned earlier, a lot of it is focused towards the younger generation, which is why they are going on to different platforms that can... that are focused more towards the youth. But then if you think of, like they already have
excellent propaganda distribution and stuff. They're not putting out as much as they used to, but now if you have a lot of these, especially when you're looking at the younger generation, they've grown up with technology, so they're learning all this quicker. So recruiting someone that's already got these technical skills has become a lot easier. So if you bring someone into an organization and you need help making propaganda, and they have these AI skills, can offer those skills to someone that may not have had as much to offer in previous years, now has these skills that can help them amplify their effort. Buddy (40:02.85) Interesting.
Heather Perez (40:02.862) And if you look at the past, like with the Islamic State, even their supporters would come up and fill the gaps. So they would have like work groups and be like, okay, we're talking to this person in, you know, Venezuela. We need all of this language. We need all of this translated into other languages. Well, they don't have to go out to their network. They can utilize AI now.
to process that information and then provide it to whoever they're talking to. Because they want to constantly engage these people in recruitment, now they have the means in order to get that propaganda transferred over into whatever they need in a quicker manner. Buddy (40:38.072) Yeah, it's interesting. you know, we get excited because of, you know, access to these technologies is being democratized and sort of decentralized, which brings down the cost and increases access. but you know, bad guys get access to that too, as well as we do. And so, it sounds like this is a problem that's here that is probably only
going to get worse in the near term, but hopefully we, we, we kind of get smarter on how to combat. combat some of these things. So Heather, as we close down this awesome conversation, what are some things that you would share with law enforcement analysts that are in the position you were in just a few years ago, and intelligence professionals as they start to think about the use of AI and the proliferation of propaganda? What are some things you would tell a young analyst today to keep in mind? Heather Perez (41:33.814) as you're building out your domain expertise, you have to understand who you're looking at, what they do, like what are they currently doing, what are their current activities, and how could AI enhance or increase those efforts? Because it may help them get into areas that they weren't previously in before.
Just as a quick example, the drug cartels have started doing a lot of fraud campaigns because it brings them in more money. So you have to look at all these different ways that their current activities can be amplified and how that might transition later. So you're not always catching up what they're doing. You can kind of like have a good guesstimation of where you think they're going to go next so you can be prepared and not constantly trying to find where they are now. When they're exploiting these models, a lot of what they're using is, you know, they're learning how to jailbreak by writing prompts certain ways. Like you need to kind
of figure out how they could potentially do that. Like how is a child predator able to go on and utilize these platforms that if you go in and type a regular prompt, will tell you that it can't provide the information, but they're able to twist it around enough to make AI generated child pornography in order to be able to kind of deal of this you need to understand how it's working and learn. Buddy (42:56.206) Well, and Heather, to think
through some of that, like we, the good guys could be using generative AI to help us think through all the things you just said. Right? Like if we're learning how to use these tools to our advantage, I don't need to think about the 10 different ways a drug lord is going to conduct operations in North Africa. Like I can actually share those ideas with AI and have AI help me start to think about it different ways. So I'm not just limited to what I know in my brain, right? mean, There's kind of a co -pilot for us too. It's not just that the bad guys get co -pilots. We get co -pilots too, potentially. that okay?
Heather Perez (43:32.852) Yes. So you have to learn how to leverage those and incorporate those into your workflow. But then at the same time, you have to remember that, you know, we're working with, you know, the ethical and privacy restrictions. They're operating with no hindrance. don't, their whole way of doing it is figuring out how to exploit it. So you need to understand how to use the tools that you have and understand their vulnerabilities as well. So it's. It's hard because the technology is advancing so quickly. It's hard to keep up with. And with
everything else that we have to do, but it's very essential. kind of like as you watch them utilize it, you can start looking at different indicators as well. I mean, it's hard. It's not now where some of the propaganda and some of the deep fakes and stuff that are coming out. It's getting more and more difficult to make those determinations if they're real or not.
Buddy (44:36.012) Yeah. Well, and I can imagine, you know, there's, there's some foundational models that everybody uses, but there's also sort of this movement with open source models for people to get access to, to, people being given the ability to now fine tune their own models. And so as more models sort of come to market and more, these models are proliferated for different purposes across different markets. it gives bad actors the ability to sort of start training their own models to do what they want it to do as opposed to, you can put all their, I guess my point is you can put all the restrictions you want on a foundational model to prevent bad actors from doing things. But if they're fine tuning and building their own models,
you can't really prevent that. You know, can't stop them really from doing that I'd imagine. Heather Perez (45:28.136) No. So at that point, it's just learning and understanding how they're utilizing their models, what type of activities they're doing, how that, like, if you come across that online or in your investigations, then it becomes just kind of, you know, sharing the methodologies and TTPs within your community. Or if it's something like for child predators and stuff, making sure you're sharing stuff with the public. So in the case of child exploitation, parents can help protect their children. But on most instances, sharing it within the law enforcement Intel community networks to make sure everybody knows what to look for, what are those indicators and signs that you're looking for. So you can learn to build countermeasures.
Buddy (46:16.034) So Heather, this has been a great conversation. know we're going to probably talk about this a lot more in the future because as you mentioned, the technology is evolving. The bad actor TTPs are evolving. And, you know, I've, I've run some polls on LinkedIn recently just to kind of get an idea of like, how, how, how often are people actually starting to use generative AI to do their job? How, you know, are we as a, as a, as the good guy, so to speak, are we in the law enforcement intelligence space? Are we investing the appropriate time and resources in learning this new technology? And you know, what I've learned so far, it's my, course, it's limited to people I've talked to in the surveys I've done, it seems like we are a little bit behind right now. People,
don't know that people are embracing this new technology as quickly as our adversaries are. And that worries me a little bit because I could see the gap sort of widening. And I do think just like talk about businesses that we've seen this in Gartner studies, we've seen this in McKinsey studies that it is very clear that businesses that adopt generative AI tools and technologies to make their businesses more efficient will create their own competitive advantages by doing so. The same thing applies in the world of misinformation
and disinformation. And I feel like it's going to take a little bit longer for the law enforcement intelligence community to adopt it as quickly as businesses do. And so that Delta, that gap between the adversaries use of this and ours, it does worry me a little bit. And so
we'll definitely be talking about this more. So thank you, Heather, for hanging out today. You have no idea how much I appreciate you coming on and sharing your experience as a law enforcement professional. You've obviously bring a lot to the table in these conversations and this topic will continue to evolve. So I look
forward to having more conversations like this. And for the folks tuning in today, just remember that as we power down today's session, remember that in the digital arena, being slow isn't just about losing the race. It's about losing everything that we guard and protect. Let's not just keep pace, but let's set the pace. Stay vigilant, stay informed, and let's turn our Intel into action. Until next time,
keep tuning in to the Wild Dog AI podcast and stay one step ahead of the pack.
2024-07-27 15:37
