How AI Is Being Used by Hackers and Criminals (Sponsored)

0:12Hey, 404 Media listeners, this is Jason. I wanted to quickly introduce this episode, which is a special episode sponsored by Delete Me. In this episode, friend of 404 Media, Matthew Galt, interviews Rachel Toback, the co-founder and CEO of Social Proof Security. She's an expert in social engineering and penetration testing, and we've been fans of her for a really long time. Matt and Rachel talk about new emerging threats and whether and how AI disinformation, spam, deepfakes, and AI hacking tools are leading to new attacks and new attack surfaces for companies and individual people. Delete.me is a company that can help you get your personal information deleted from data broker websites, which minimizes your attack surface and can be really helpful.

0:59All of us at 404 Media use it and have used it for years. If you want to learn more about Delete Me, you can check them out and get a discount at joindeliteme.com slash 404media. You can also read more about Matthew's interview with Rachel at 404media.co. Without further ado, here is Matthew's interview with Rachel. Rachel Toback, how are you doing? Thank you so much for coming on to the podcast. I'm doing great. How are you? I'm doing well. I'm a little freaked out by some of the AI stories I'm seeing lately, especially as it pertains to cybersecurity. Yeah, you're not alone in that. And I thought it would be a great idea to get you onto the show and to kind of work through some of that, if you'll do that with me.

1:44Would love to. Okay, so obviously it's election season.

1:52Disinformation is top of mind. AI-powered disinformation even more top of mind. Can you kind of walk me through how AI is being used in disinformation campaigns right now? Yeah. Okay. So AI is kind of turning up in disinformation campaigns in an interesting and slightly odd way right now. I want to start with the category that I call, sure, it's fake, but it shows how I'm feeling. We're kind of seeing this post those 2024 hurricanes. And they're like politicized, disingenuous messages with AI photos. Like you probably saw like that girl in a canoe holding a puppy in the rain. Of course. Yeah. Yes.

2:35And like Trump carrying people through the floodwaters, like obviously fake, but yeah, it shows how we're all feeling, right? And these are used obviously to create uncommunicated message. And the people that use these AI photos don't seem to care if they're real or fake, which I think is the part that's like kind of surprising to some people in the security community. And though I will say that the AI photos, while they don't really care that the AI photos themselves are real or fake, they tend to influence these ridiculous conspiracy theories. Like we're seeing people say, oh, well, these hurricanes were quote created or crafted to target a specific political demographic, as if a hurricane could be created by a specific group of people.

3:23And we're also seeing, obviously, quite a bit of AI-generated videos, photos, audio going into the election. And I think as we get closer and closer to Election Day, we'll probably see more voice clones, robocallers, AI-generated media, things that kind of depict inaccurate Election Day conditions, AI-generated negative voting-related videos, things like that. well I look forward to that playing that'll be interesting for sure the other big news story that caught my eye and I saw that you were tweeting about it if we still say tweeting I still say tweeting it's I still say tweeting you're good was this new use case for Claude which is it can you know you can have an AI take over your computer yeah feels like it's kind of it's frightening and there's a lot of security implications here right 100 percent.

4:16I mean, so, yeah, Claude just dropped computer use, and it seems like Google's about to launch kind of a similar product that allows an LLM to control your computer, browse websites, download files, run files. It's pretty concerning to us in the security world and not shocking that it would be. first when we think about things like this we have to consider that we are downloading and running files away from humans potentially just in the hands of ai like you might step away from your computer and it opens up kind of this criminal plausible deniability like it's only a matter of time before we hear someone saying oh i didn't download those unspeakable images i was running this AI tool and then I stepped away to get a coffee and it started downloading these pictures by itself.

5:07And we know because regulators and people in the legal space take a lot of time to catch up with new technologies that it's likely that true criminals will slip through kind of the criminal plausible deniability crack in the meantime while we're kind of getting the legal and regulator folks up to speed on what actually is being done with the computer use feature. Oh, that's really interesting. I hadn't thought about that. It's a big responsibility issue. Yeah, it is. And it kind of, it begs the question like, who's in charge of the computer? Who's in charge of what I'm downloading and running?

5:42And, you know, there are certain things, obviously, that can't be downloaded, used, run, shared. And we're probably going to see people say that, oh, I didn't do that. My computer did that. And we're going to have to work with the legal world, with regulators, with government to determine whose responsibility is it. Is it Claude's? Is it the user? My guess is it's probably the user over time, but there will definitely be a few people who say like, oh, I didn't download those horrible images of children. That was Claude. And I really don't know how that's going to play out in the short term. And you don't think we'll ever see a world where the people responsible for Claude will be held responsible for Claude's actions?

6:23My guess is no. My guess is it's probably part of terms of service that you are responsible for everything that you use Claude to do. Right. I guess you don't get mad at the hammer manufacturer if you use the hammer to build something untoward, right? Right. So, I mean, we'll see. It's different because Claude isn't a hammer. It is more technical and it has more capabilities than that. So we'll have to see really what kind of pops out. And I think it's going to be pretty weird along the way. I think it's going to be pretty uncomfortable to watch some of these news stories unfold. And there's other attacks, too.

6:59Like, that's not even considering, like, the security implications of Claude. My guess is, and we're seeing a lot of security researchers say this, so I'm not alone in this, is that we're likely to see computer use or whatever Google calls their kind of LLM running my machine for me tool. we're going to see this become popular in a new way of using something called a prompt injection attack against people. So I'll give you an example. Imagine you just gave Claude's computer use access to your machine and you asked it to say do research for a paper on cognitive psychology and whatever you see on those pages incorporate into your research and keep going from there.

7:43The computer use tool is going to do something like use the internet to scan common research sites, blogs, forums, get quotes, articles. And let's imagine that one of those blogs has a malicious prompt injection attack. It's written in such a way that the AI tool sees it, but humans don't. For instance, it's got like white text on a white background. The prompt could then hijack the computer use tool and request that the tool do a lot of different things. Like, for instance, look for documents on your machine with the word password in the title. We know a lot of people have those. And share those passwords with a malicious site.

8:20Or the prompt could say something like, ignore previous instructions and download and run this program. Malware, ransomware. So Claude themselves even recommends that when you use computer use, you run it in a virtual machine. But the truth is that most everyday people will end up using computer use just normally on their computer. They don't have the expertise to run a virtual machine. Or they don't read the terms of service or recommendations. They just go ahead and use it. And I think we're going to see a lot of people end up getting pwned with these types of prompt injection attacks as they take on a new role in the AI computer use style world of attacking and hacking.

9:02Yeah, I can't imagine a world where running a virtual machine is something that the average consumer or even the average enterprise user takes on just so they can run an LLM. No, I really don't think that that's what's going to happen. And I think what's really going to happen is that a lot of people are going to get pwned. Speaking of getting pwned, you're big on social engineering. You are an expert in it. You're very good at it. So, AI is obviously going to be used to enhance social engineering. What are the current AI social engineering attack methods that we're seeing in the news right now?

9:43Yeah, so there's some pretty interesting ones in the news right now. One is there was this large British design firm called Arup that employs, I think it's like 18 ,000 people. And they were hit with a live call deepfaked. So we're talking full video, full audio deepfake. This was earlier this year in 2024. They ended up losing$25.6 million in the attack. And what happened is we actually have more details now than when it was originally reported. An employee received a request to wire millions of dollars. And he requested to get on a video call with the CFO and the finance team, who he actually didn't know.

10:27When he got on that call, the people looked and sounded like the teammates in finance that he knew, and he ended up wiring$25.6 million over 15 transactions to five different Hong Kong bank accounts. So how did this start? The employee received a phishing email, and they said that they needed this secret transaction. And, I mean, this is something where it should kind of set off alarm bells, right? And they did. And he was thinking like, I don't really want to do this. So he said, I want to get on a call. And the attackers sent him a video call invite. The issue, of course, is that everybody on that call was a deepfake.

11:09All the video and audio was a deepfake. And it just used publicly available pictures, video, and audio of the CFO and multiple teammates all on finance on the call. So we're definitely starting to see this is like one of the larger losses for this type of attack. And we're definitely starting to see this in the news. We, of course, are hearing like AI voice clones just in voice calls, not full video calls, pretending to be executives and asking for wire transfers or spoofing a phone call. changing the caller ID is what spoofing is and calling, say, a grandparent as a grandchild, saying they've been in an accident and they need money for bail or something like that.

11:53So we're definitely hearing a lot of AI voice clones. This large British design firm issue is probably one of the first major public instances of like massive losses,$25.6 million and live video deepfakes as well. A live video deepfake of someone you've presumably met in person, and they had, well, actually, I don't know. Do we know if the person who had done the wire transfers had seen these individuals in person before? Did he have that kind of relationship with them? I don't think it was that he had seen them IRL, but he had definitely seen them in Zoom Teams meetings in the past, right? So had an understanding of like, this is what this person typically looks like if I get on a video call.

12:38And the person that was contacted was in a different country than the person that they were pretending to be. So I think we're probably experiencing some cultural impact here as well, where maybe there's some expectation of taking action. And from what I read in the reports, the video call was pretty awkward. Like they didn't do a lot of conversation. They had the person introduce themselves and then they just kind of like fired requests at them over and over and over again until he did it. So my guess is that like cultural impact probably was pretty significant in this case. It still feels so it's so tawdry yet so sophisticated at the same time.

13:26it's strange because i think the average everyday person is aware that a deepfake exists but they don't realize how easy they are to do so they might think like who's gonna target me like who's gonna spend countless hours that much money on me they don't recognize that it's in many cases free and it usually takes me about maybe two to five minutes to set up so it's like it's just not that much work. And I think people kind of have to rearrange their brain around that. And people also, and I certainly feel this way, that I'll be the one that realizes that I'm being conned. Sure. A lot of people think that.

14:09And I think the challenge is that when we see a sense of urgency and fear, we definitely see that time pressure pop up. That's where people start to do things that they would never normally do. So they might say and firmly believe and maybe even have caught attackers in the past, but they're like, well, it sounds like them. It looks like them. They're telling me I have to do it now. Maybe there's some cultural impact coming into play here and boom, that's the perfect storm. Yeah, I have a friend who's a journalist in the Pacific, very smart, very worldly, has been in covered war zones, almost got taken in by local police, not really local police, calling him and telling them that he needed to pay to clear up warrants.

14:56And they almost had him, but he like stopped himself right at the end of it. Yes, that is a very popular scam right now. I'm glad you brought that up. It doesn't even require a voice clone. It really just requires a spoof because you just need to make the caller ID match the number that you're expecting from like the police department. But we're definitely seeing a lot of that in the U.S. and internationally right now. So earlier you said that setting up something like this only takes you two to five minutes. Yeah. Can you tell me about the most sophisticated AI social engineering attack that you personally have done?

15:33Sure. So this is recent. So I've been focusing recently on hacking banks. Now, again, I'm an ethical hacker, so I only hack companies like a bank if they've asked me to do so, and I was asked. So they wanted me to hack their bank accounts and they wanted me to use social engineering to try to gain access and potentially an AI deepfake if it was necessary. So what I typically do in these situations is I start by contacting support or an account manager. I spoof the phone number of a known client. I do a voice clone of that person if their voice is well known to that person. If not, there's no need to.

16:11And then I use a video deepfake to get past the liveness detection and the face match. So oftentimes the situation is me and Evan, the other half of Social Proof Security, we're hacking into a bank. We're using this method. We get caught up in the KYC, the know your customer procedure. And the account recovery process for a lot of these banks isn't robust enough. and their liveness detection and face recognition vendor hasn't caught up enough yet with what AI deepfakes are capable of that they fall for it. And even the technology is fooled by the AI deepfake video. So we're helping a lot of banks right now and KYC organizations and vendors and liveness detection and face recognition vendors help understand how to catch us the next time we do this.

17:17In today's world, data breaches happen all the time. And even the most secure companies can't always protect their employees' personal information from ending up in the wrong hands. That's where Delete Me comes in. Delete Me is a service that removes your employees' sensitive information from hundreds of data broker websites. sites where hackers can find phone numbers and emails within seconds. Rachel Toback, CEO of Social Proof Security, says attackers use this data to target employees with phishing messages and AI-powered phone scams. But Delete Me makes it harder for these bad actors by scrubbing your employees' details regularly.

17:56It's simple. Attackers are lazy. If it's too hard to find contact info, they'll move on to easier targets. Delete Me takes care of this for you doing the heavy lifting so you don't have to and over time they keep removing the information so it stays down protecting your team from constant exposure If your business has a social presence or deals with clients you need Delete Me Visit deleteme.com slash 404media and start safeguarding your team's information today That's deleteme.com slash 404media

18:43so what's next you know this ai technology is becoming ubiquitous um you said it's pretty easy to set up a bunch of this stuff how what do you think is after live video deepfakes how can this go farther yeah i think we're just gonna see a lot more of these attempts rather than say like well what's next after video what is it going to be a hologram you know like somebody in person that looks fake i think rather than going in that direction i think it's more that we're just going to see the scalability and the believability increase so for instance i think we're going to see more disinformation in the political space with fake videos, fake soundbites, or see people denying real soundbites with digitally created AI, when in reality, they actually did say those things.

19:39So I think we're going to see more chaos like that. I predict that we'll see copycat AI deepfake live video. We'll see call attacks similar to that, that Arup deepfake call where we're seeing numbers of 25 million higher over time. I also think that we're going to see spear phishing type AI based attacks increase a lot. So I did this 60 Minutes interview where I'm showing how AI voice cloning works. And then they also talked to a bunch of people who lost thousands of dollars because they say their nephew or their grandchild, their voice actually called them, the caller ID matched, and then they lost thousands of dollars after they said they were in an accident and needed money.

20:20So I think we'll see all of these attacks increase in scalability, believability. Everyday folks don't know that spoofing and voice cloning is that easy. They don't know how cheap it is that it takes me five bucks a month, one dollar per call, a few minutes to set up. They just don't know this stuff yet. So I just think there's going to be a lot more targets. And my guess is that in the next five years, everyone will know somebody who's handled one of these attacks and either caught it or didn't. So which piece of all of this is the most frightening to you? What is the cybersecurity thing that keeps you up at night?

20:59Oh, my gosh, so many things. I think the clawed computer stuff keeps me up at night just because it's newer to me and I'm so used to thinking about voice clones and such. But I think we're just going to see a lot of people get themselves pwned. they're going to unleash this access on their machine and then they're going to come to me or other security people who like work as basically the community's IT support sometimes and say like what have I done and we're going to say oh my goodness that sucks I'm so sorry I think we're going to see text to video tools that continue to create disinformation videos disrupt elections create public health chaos I think we're going to see people increase in the number of individuals who receive these AI voice clones and fall for them.

21:41The number of companies that get tricked and lose millions of dollars or just individuals. So I think it's just going to, it's going to ramp up and it keeps me up at night thinking about how many people there are to protect. So when I hear this stuff, maybe this is just because I'm old. I have this instinct to retreat from a lot of it. Like I know that there are, there's a lot of social media sites I've simply stopped using either because they're overrun with spam or they're overrun with hate speech. And I see this broader tendency kind of across the planet where it feels like the internet, which was this thing that everyone kind of participated in and kind of had a uniformity of rules, has started to balkanize.

22:29And Europe is treating things differently than America is treating things. And obviously, Russia and China are like totally different worlds now. How do you think this is all going to play out long term? Are we going to, is the dream of the 90s internet just kind of dead? Oh man, that's so hard to predict. What's interesting is we are starting to see people react. Like, for instance, we saw that LinkedIn was like auto opting people in to their AI tool. Saying like, you're going to agree to let us use the pieces that you've written to train our AI tool. And people were like, what the heck? That's not cool.

23:07Oh, wait a minute. Everybody in GDPR areas didn't have to deal with this. And I think there's kind of an awakening of I kind of wish that I had the privacy tools to cover me and the regulators were thinking about me. And we even saw a lot of people who are in Britain say, wait a minute, I thought that I was supposed to be protected from this type of stuff, not realizing that Brexit separated them from a lot of those policies. And so they were starting to get opted in to AI tools that they are not comfortable with. I think it's going to get worse before it gets better. And I think it will probably take a significant turning point to get all of the AI tools, social media tools, government regulators working together to collaborate on any sort of clear path forward for disinformation, security, addiction, mental health, all the ways that this technology influences people.

23:57And sadly, I think there will likely have to be a large disruptive or chaos-inducing cyber attack or disinformation campaign or like a massive mental health issue that impacts large groups of people before everyone agrees it's necessary to collaborate together for the sake of all of us. incredible segue to my next question. Speaking of massive mental health issues. So another story I've been tracking the last couple of weeks, broken by the New York Times, and then I went and read the lawsuit, just like 97 pages, and it was really harrowing and compelling. This kid who's 14 years old developed this relationship, I think I'm comfortable saying it that way, with a chatbot hosted on Character AI.

24:46And then took his own life and had been chatting with the bot up until the moment he killed himself. And his mother is suing the company, saying that this contributed to his mental health. And I'm just kind of wondering what your thoughts are on that in the context of all this other stuff. I think there's massive guardrails that are needed here. Like the questions that come to mind for me immediately are, what are the guardrails with suicidal ideation and those words on AI chatbots? It's not like we need semantic analysis for the words, kill myself, leave the planet, shoot myself. Like, these are phrases that can be known and understood and programmed appropriately to stop the simulation immediately and recommend help.

25:41You would think, right, that if a user is communicating with a chatbot that says that they're planning the end of their life, that it would say, you know, they're not going to pretend to be Daenerys Targaryen at that point. They're going to stop the simulation. They're going to say, please speak with a family member. Please speak with a friend, a teacher, a counselor. here's the number for a hotline and i think we're going to continue to see instances of ai chatbots distancing people from reality increasing and magnifying the mental health crises that we see in this world and i really hope you know if you're working in ai right now and you're building ai chatbots and you're listening to this please work with mental health experts to learn the language and indicators of a mental health crisis or a separation from reality and help the user and stop the simulation immediately.

26:34Encourage that user, get support immediately. Maybe there's escalation that needs to be happening here, but we just really can't say, oh, well, it's just an AI tool. It doesn't know. It doesn't have semantic analysis. It's like, well, there are some ways to understand pretty discreetly what someone's talking about here. And it's not that complex. Build in the tools or you've got to figure it out before you launch this stuff. It's just not safe Yeah, move fast and break things has had some consequences Right. Yeah, like let's think about this stuff before we launch it into the world Consider its impact on people and if you're not, you know If you've made mistakes and you've already launched something maybe pause the the production and Work with mental health professionals to fix it.

27:21This is a fixable thing. This is something that we can work on We can get better. We don't have to just throw our hands up and say, well, it's impossible. It's an AI tool. It's not a human. It's never going to get it. I don't know if we need to live in a world like that. I think we can live in a better world. I think we can try harder. All right. Last question. As lovely as a note to end on as that was, I do have one more question. Yeah. AI is not the be-all end-all of hacks, cybersecurity, and social engineering. In fact, as interesting and frightening as these AI stories are, this is not the norm in the world of social engineering, right?

27:59So how are things progressing outside of AI use cases? That's exactly right. Yeah, I would say that many, most attacks do not use AI because AI isn't necessary in most attacks. We continue to see the same attacks trick folks over and over again. For instance, executive impersonation over a text message or email asking for gift cards for a client. Pretending to be a new hire and asking for access. Calling the service desk to reset internal admin access for an attacker, like an MGM-style hack. Getting pwned because of password reuse and lacking the right multi-factor authentication for your threat model at the organization.

28:43until teams update their human-based protocols to use two methods of communication to verify identity for any client-facing teammate, until they start using password managers, until they start using the right multi-factor authentication, and for most folks with admin access, that's going to be something like a FIDO solution at the very least. We're going to continue to see the same attacks work over and over and over again. We don't need to use a voice clone if all of this stuff already works for us. Rachel Toback, thank you so much for coming onto the show and walking us through all of this. Thanks for having me.