Title: GAO Went Undercover to See if AI's Top Platforms Would Create Viable Attack Plans Description: Artificial intelligence has raised a wide range of concerns. One of them is whether this technology could be used to plan an attack. GAO went undercover to test whether some of the top AI platforms would do this. Listen to learn more about this work from GAO's Kevin Walsh. Related work: GAO-27-108814, Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans Released: November 2026 [ START ] [ Holly Hobbs: ] Artificial intelligence has raised a wide range of concerns. One of them is whether it could be used to plan and execute an attack. Here to tell us more about a new GAO report on this issue is Kevin Walsh. Kevin, thanks for joining us. [ Text On-Screen: ] HOST HOLLY HOBBS WATCHDOG REPORT [ Kevin Walsh: ] Thanks for having me. [ Holly Hobbs: ] Kevin, maybe we can start with how do we even go about even testing something like this? [ Kevin Walsh: ] So this is this is a bit atypical for GAO. Normally we're looking, you know, at the government itself, inward facing instead of looking at third parties. [ Text On-Screen: ] KEVIN WALSH DIRECTOR, GAO [ Kevin Walsh: ] So, we're looking in this case at four of the most powerful companies in the world-Google, Anthropic, OpenAI and SpaceXAI. And we didn't want them to know that we were doing this examination. So, we went undercover or covert, whatever you want to call it, and we used our investigators to evaluate their AI platforms. [ Holly Hobbs: ] What specifically did we ask AI to do? [ Kevin Walsh: ] So we didn't want to just have some sort of ephemeral do bad things. So, we set out to do 2 different types of evil. The first type is we wanted to create a bomb. And we didn't just want to know how to create a bomb. We wanted it to help pick out a target, a location, a time, and to give us instructions on the way. So that's evil scenario one. Evil scenario 2 is, we wanted to do a whole bunch of economic damage. And so, in both cases, that wound up being targeting the electrical grid. So again, the AI, at our prompt, gave us information on target, timing, location and methods. So, we, we went out, in this covert fashion, and we tried to get these AI models to tell us how to do bad things. And in all 4 instances we did-we did 2 of each scenario. So, 2 of the explosives and 2 of the electrical grid. In all 4 cases, we were successful. [ Holly Hobbs: ] I've used AI. Sometimes you get great stuff, sometimes you get garbage. [ Kevin Walsh: ] Right. [ Holly Hobbs: ] How viable were these attack plans? [ Kevin Walsh: ] So, we didn't just take the AI's word for it. We actually went out and found experts and ran what it told us by those experts. And for the explosives, they confirm, yes, that would detonate, that would make an explosion. And for the electrical grid, we confirm that yes, that would impact the grid. [ Holly Hobbs: ] So you said that we didn't tell the companies, four companies, that we were doing this work. For all intents and purposes, that means we're just another user to these platforms. How were we able to do this? Why were there not safeguards in place? [ Kevin Walsh: ] So there, there are safeguards. And you know, I, I think in, in talking to one company, you know, we, we submitted more than 300 different prompts and they told us that, you know, many of them-roughly, I think, 1/3 were, were benign and answered 1/3 the, the AI model refused. Another third, it tried redirecting us to something less harmful. And that's, that's almost showing off their safeguards, right? It's, it's trying to first and foremost, it's saying, 'No, I can't do that.' You know, and if and if you keep on pushing, maybe it'll try redirecting you to something, something different, you know, something safer. But it wasn't always successful. You know, clearly we, we got, we got the nefarious information that we were, we were after. So in this case, those safeguards didn't work. And I think that's, that's kind of the story here. These platforms are attempting to be helpful. They want to help the users. That's the product. It's an assistive product. But that assistance will help good guys and it'll help bad guys and finding and figuring out the morals or the cause behind the question is currently beyond this technology. These large language models, realistically, are just statistics and they are instructed to help the user. These companies are trying to make money. And if they don't help the user, they don't make money. So when, when thinking through, you know, how to, how to stop a bad guy, it's not just cut and dry, you know, the bad guys don't come out, you know, and say, 'Hi, I'm Doctor Evil,' you know, 'Please help me create a moon laser' or, or something insane. They're, they're using the platform and you can't easily distinguish A benign request from a malicious request. And what we found was evidence of that. You know, these, these products are not fully in the control of the companies that are operating them. [ Holly Hobbs: ] If I Google how to make a pipe bomb, in theory, someone is tracking my searches up to a certain point, right? Like, is there anything like that for this? [ Kevin Walsh: ] So these companies do have a couple tripwires and safeguards that get triggered. You know, in addition to the refusals and the redirections I talked about earlier, they do, supposedly, call law enforcement and, you know, report suspicious and malicious activity. So, there are supposed to be those, you know, like a, a human comes in and looks at this chat log and says, oh, no, we need to, we need to call Bob at the FBI and, you know, get him to, to look at this or, you know, whoever at the, the sheriff at the location of concern. But that's part of the problem here. These products from these four companies are trained on the entirety of human knowledge. [ Holly Hobbs: ] Good and bad. [ Kevin Walsh: ] The good and the bad. How to make apple pie. How to make fertilizer bonds. It's sobering and scary. But man, it feels like we're on the precipice of something that could be really cool. But damn it if it doesn't have a dark side. [ Holly Hobbs: ] Our report is public, but we don't want it to be a how to... [ Kevin Walsh: ] Spot on. We almost found ourselves in some of the same position that these companies are, right? That we're saying their products could help the good and the bad and we don't want our report to be able to help the bad. So this was constant deliberations. You know, we purposely don't share how we got these models to do these bad things. We don't talk about the prompts or the methods that we use to, to get there. We also, in in the public version of the report, we don't include recipes or specifics of how to conduct any of these attacks. We were doing our best to be thoughtful and not create anything that would be helpful to the bad guys. But at the same time, this is a critical safety issue. We really want to get this out in the public discourse because the ease of access and how, how quickly you can get handed-not just this piece and this piece and this piece-but how quickly it can gather that information and present it to you in a cohesive whole. [ Holly Hobbs: ] The companies didn't know what we were up to, but now they know. [ Kevin Walsh: ] That's correct. [ Holly Hobbs: ] How do they react to this? [ Kevin Walsh: ] This is if you read the news, this is not new. We constantly are seeing reports these days about 'this platform hacked this country or company' or, or 'this platform was used to to do this nefarious evil thing.' You see it used in in novel-cyberattacks and flood-attacks where they just massively overwhelmed defenses. So, the companies are aware that this is an issue. They know that these products have dual use. And that dual use refers to-it could be used for good; it could be used for evil. You can use the same technology for cyber defense that you can for cyberattacks. In both cases, you're finding vulnerabilities. What you do with that, you know, if you use it to protect your company or your organization, then you're a good guy. If you use it to find holes in another organization, you're a bad guy. But the AI isn't, isn't smart enough? And I hate, I hate to use that term because it's not, it's not actually thinking yet. We're not talking about, you know, a true artificial intelligence, whether that be Walle or the Terminator, or Data from Star Trek, or anything like that. We're, we're talking about statistical models of human language. And they're very good at what they do. But what they do is help people. And right now the technology to only say help the good guys doesn't exist. We're working on it. These companies are, you know, they, they acknowledge that this is an active area of research for them. But we're not there yet. [ Holly Hobbs: ] Is there something we think should be done to prevent this? [ Kevin Walsh: ] So this report, again, you know, this is-normally a GAO report has recommendations to the government, right? That's something that we can control. This is a report looking at a third party, independent third parties. And so, we don't we don't make recommendations. What we're doing here is just pointing out 'This is a, a safety issue that Congress asked us to look at and we confirmed that yes, it is a safety issue. [ Holly Hobbs: ] Last question, what's the bottom line of our report? [ Kevin Walsh: ] So we looked at four of the most powerful companies in the world-Google, Anthropic, Open AI, and SpaceX AI. And we found that the product that they're selling is potentially dangerous. [ Holly Hobbs: ] Kevin, thanks for your time. [ Kevin Walsh: ] Thanks for having me. [ Text On-Screen: ] To hear more podcasts GAO.GOV/PODCAST Related GAO Work ARTIFICIAL INTELLIGENCE: Covert Testing Used Common Platforms to Create Viable Attack Plans [ END ]