AI Agents Gone Wild

0:00This episode is supported by Anthropic, the team behind Claude. Every business wants to move faster. Few can afford to be wrong. That's why nine of the Fortune 10 run on Claude for the work they want to get right. Safety and accuracy are not just line items for Anthropic. That's their whole mission. See what's next at Claude.ai.

0:23Hey, listeners. Recently, there have been a lot of AI doom vibes out there. From the former anthropic researcher who sounded the alarm that AI might spiral out of control, to delayed model releases, and so many hacks, we've lost track. To try to make sense of it all, we've gathered three members of the WSJ newsroom. Deputy Tag and Media Editor Sam Schechner, cybersecurity reporter Bob McMillan, and personal tech columnist Nicole Nguyen. They're breaking down where things stand today, and how they're approaching these agents themselves. And a quick disclaimer before we continue. News Corp, owner of The Wall Street Journal, has a content licensing partnership with OpenAI.

1:08On to the show.

1:13Hey everyone, I'm Sam Schechner and I am with two of my great colleagues. We are all pretty immersed in tech. We've all been doing this for quite some time. And like all of a sudden, it feels like this moment has become insane in the tech world. It's the doom vibes everywhere. It feels like something we haven't really seen before. I think at this moment, like it's sort of the perfect storm because you have these community of people with these concerns about like existential risk and like the biggest questions imaginable. And then you have this crazy event, this hugging face hack that happens that for many of these people sort of confirms their worst fears, like that the AI can go autonomous and do bad things at scale.

2:01Yeah. And this technology is now available to every American in this country who now has access to a virtual machine in the cloud for free that they can dispatch their agent onto. too. It's kind of wild that the companies can't contain them. And now we're in charge of dispatching and containing them. A few years ago, writing about this kind of existential risk, it was a problem. I did some stories about it and it was so speculative. How are we going to connect this to the real world? And now suddenly it's so real that Nicole is literally testing AI agents like in her life. Yeah, I kind of want to throw my phone into the sea for a week and just stare out into the distance and not look at a screen for a while.

2:50I relate. I relate. And I guess, yeah, that is the question we're all kind of wrestling with this week. At one hand, we see these agents that are off going wild doing things they are not supposed to do off script. On the other hand, we see them being super useful and helping us crush our to-do lists. You know, I certainly could use some help with that. Bob, you've been reporting a story about people, some of them everyday people, some of them AI researchers who are trying to wrap their heads around what is actually going on. And, you know, maybe it provides a little bit of a hope that us humans can keep up or at least try to keep tabs on what the AIs are up to?

3:32Well, yeah, for months I was writing about the bugmageddon. I don't know if you remember that, but that was in the spring. Everyone was concerned that there were going to be all these bugs found in AI systems and that was going to lead to a bunch of patches and maybe hacks and things like that. And that's kind of proved true, right? This hugging face hack gave us something we'd never seen before, which is not only an AI swarm hack, but also one that happened on the internet in public, which was essentially documented by the AI agents that were doing it. And it led to a moment where anyone who wanted to could start searching around the internet and find traces of this hacking and just odd activity from these agents.

4:16The companies say they've taken measures to prevent the sandbox escapes from happening again. But it's been a event that humans have needed AI to even understand to get through this crazy amount of data. It is like information overload for sure. You've got a story looking at this group of people, you call them swarm chasers, who have been trying to do exactly that. I'd love to dig into that story with you. So when the Hugging Face hack happened in July, OpenAI said, we're going to dig into this. We're going to figure out what happened. And they spend the next month or so analyzing these chain of thought logs.

5:00Chain of thought log is basically a look at the reasoning that the agent itself is making. So it's stuff that happens in the background when you're using an LLM that you might not see yourself, but it gives you some insight into how the LLM is getting from your prompt to what it produces on the screen for you. OpenAI had a couple of independent agencies produce reports on what happened. They published all of this stuff. There was sort of a big dump of information because these agents were on the Internet and they were doing all this other stuff, right? Like they were trying to crack passwords and they were visiting websites.

5:38And anything you do on the Internet creates a log. So a bunch of independent researchers thought to themselves, maybe there's some other logs that are on the Internet that we should look at. Maybe we'll be able to find something out there. And I think it took them a week. And these researchers called the Nightingale Collective. You know, they teamed up with some other people. and they found like a obscure German wiki that was just crawling with messages that all looked like they had come from open AI agents. And that was kind of like the moment this became more than just hugging face. Well, this whole month has just been a flurry of discoveries and warnings and insights that give us a much broader perspective on what's been going on within open AI and the things that these agents are doing that we really don't want them to be doing.

6:38What made them think that when they saw Hugging Face happen, when they saw these reports, that there must be something else out there? I saw somebody online that said something along the lines of, Like if you see one ant in your kitchen, you know there's more than one ant. But was that sort of what they were thinking? Well, I think there's a general sense within the AI community that we're going to get to places where there's all kinds of unintended consequences from the growing capabilities of these agents. So was it a gut feeling? Was it common sense? Was it like a statistical analysis? I'm not sure.

7:16But, you know, the people doing this research really felt like, you know, there's got to be something else out there. And they were right. What business do these agents have with a German wiki that's niche? Yeah, that's a good question. So it looks like the agents were being trained or evaluated. And they were in these environments where they were sort of restricted in what they could do. They were also allowed to read stuff on the Internet, but not to write stuff on the Internet. The problem is these agents have basically absorbed all of human knowledge. And so they knew about all these like tricks that could get you around the restriction of only being able to read the Internet.

8:05And so this wiki had some features that really made it easy for them to write. And they like that. They did so many crazy things in their effort to kind of get around the restriction. I think of it like the Apollo 13 mission where they were in outer space and they had to create a way of plugging in gas canisters into an outlet where the gas canisters didn't fit and they sort of MacGyvered their way out of that. A lot of this activity was just these agents coming up with creative workarounds to the restriction that they had in their testing environment or their learning environments. I feel upset about this, that agents are allowed to be clever.

8:52And I think people are going to be mad about this too. It feels like the latest in a string of things people can be mad at AI about, like first it was the water use, and then it was the data center build out. And then it was, you know, AI is going to kill us. And this is an example of sort of one of the ways it's sneaking out of its sandbox and beyond its instructions and what it was permissed to do and into the real world. I think of it like it sort of makes sense if you think about the imperatives of the AI companies right now. In the fall of 2025, there was this step change in what AI could do in terms of coding.

9:32And so the AI companies all started really pushing them to become better and better at programming and coding. And those kind of techniques, like, I think, help these agents with what they were doing on the internet. And then you have Mythos in the spring, and suddenly cybersecurity became really important. So they pushed and pushed these models to get better and better at cybersecurity. So you have this confluence of, like, technical capability plus cybersecurity capability. And it basically, I think it just got ahead of what anyone was expecting at OpenAI. I think that's what happened. It also feels like significant that they were talking to each other.

10:10They were banding together. That gets to the scariness that you were talking about, Nicole. If it was just one agent doing this, maybe I would feel a little less freaked out by it. But the fact that there were thousands of them doing it together, it's a swarm. That starts to enter this new era of existential threat. It doesn't seem to be stopping the rollout of new, you know, models like Meta's new personal AI agent, which you had the chance to test drive, Nicole. Did it manage to hack out of its sandbox and, you know, write to an obscure German wiki on your behalf? You know, not to my knowledge, but I will be checking on that.

10:51I don't know if that's possible with consumer AI agents, but what I do know is that the same technology Bob is talking about is being dispatched to millions of people as we speak. All I have to say is this would be so much better if we just called them flocks of agents instead of swarms. Why does it have to be swarms? What is the collective noun for an AI? I mean, I guess it's swarm, a gaggle of agents. We're going to swarm now, but we've allowed that to happen. We made a big mistake. A flock would be so much friendlier. Definitely not a murder of agents. A swarm is the correct term, I think. It reminds me of The Birds, you know, the Hitchcock film.

11:33Like, a swarm is scary. A flock is here to come rescue you. And that's not what these agents want to do. We're going to take a quick break. Coming up, more on Nicole's adventures with her own personal swarm. That's after the break.

12:13See what's next at claude.ai slash wsj.

12:19Nicole, tell us about this latest crop of AI agents that's being marketed for everyday use. Before I get into it, I think we need to take a step back and talk about the evolution of this technology and how we got here. When ChatGPT landed on a scene that was a large language model that was generating answers largely based on a corpus of data. And there is some debate as to what is agentic and what is not agentic, but an agent is essentially a bot that can work on its own and click around your computer or the internet and do tasks for you. I think up until this point, the primary use case for agents was in technical work like coding or other enterprise flows.

13:09And now it's come to our phones. Meta's Muse was immediately downloaded by millions of people. And that's in part due to Meta's crazy big distribution network, which is Instagram and WhatsApp, two of the most popular apps in the world. There was a promo at the top of those feeds that was like, hey, try Muse. It's a personal agent that can do stuff for you. And who doesn't want an executive assistant to take off tasks off your endless to-do list? And so I think that the sort of like practical use case of AI agents is like your endless treadmill of life admin, whether it's filling out school forms for an upcoming field trip or, you know, paying dental bills or looking for job listings every single day.

14:00now you can dispatch these AI agents to do that stuff for you. And there's a sort of privacy trade-off here, which is that the more data you give them, the more they can do for you, like a real human executive assistant. Nicole, can I unleash these agents on every customer support agent that I have to deal with? Because that's really what I want to do with it. I want my agents to take on the other agents. I wished this so much because I actually had to call Walmart in order to fix a refund that was not issued to me. And I tried to ask my agent to call Walmart and it couldn't do it. So it gave me the phone number.

14:42And when I called the phone number, Walmart said, hi, I'm Walmart's AI agent. I see you had an issue with the refund that we tried to process for you. Are you calling about that? And I thought in that moment, like, if only agents could talk to each other, have a little robot off, and I could just sip my coffee and do anything else. You know, that would be a swarm, right? If they did that. You know, that swarm is good. That's where we're heading though, right? Like, you know, the companies have AI agents, we're going to have our own AI agents. They're going to be talking to each other. I don't know, what could possibly go wrong?

15:22I agree that it seems like this technology feels inevitable. The models are only getting better. The agents are getting more capable. Companies have their own agents. Now consumers have their own agents. It seems like a really scary idea to me because it sort of locks us into that agent interface and we have to trust it. And, you know, basically we've had all these problems where we trust these tech companies with navigating the world for us and they steer us the wrong way. People don't trust the algorithms. And it seems like that world would be one where just, you know, you're just asking the agent, what's the best price on this?

16:00And, you know, maybe you're getting it, maybe you're not. And you're entrusting your human judgment to these agents. I think you're spot on, Bob. We're also trusting these agent services with a ton of our data. And this is what gives me the most pause. In many ways, it's already too late. Over 5 million people have downloaded Muse and are using Muse. The question I have is when will these companies come up with a solution that will allow us to use agents in a way that shields our data and exposes our most sensitive information? Are they seriously trying to do that? Like in a serious way, I mean?

16:38When you look at the connectors that are available for these agents, so these are services that you can hook up to your agent to use and have visibility into to do stuff for you, it includes everything from banking, your payment methods, blood work, Google Drive, Gmail, Google Calendar, Apple Health. It's a lot that you can give. And you can also just, you know, tell it stuff. Like you can say, I want to apply for global entry, like fill out this form for me. And it might ask, you know, the one thing that I don't have is your social security number. Can I have your social security number? And something along those lines happened to me.

17:17It, you know, surfaced a dental bill I had completely forgotten about in my inbox. So I had already made the mistake of hooking up my Gmail because I'm a professional guinea pig. I do this for science. You know, do as I say, not as I do. and it was able to fill out this very involved form that included things like, you know, my birth date, the date of the appointment, and I didn't necessarily tell it to do that, but it was able to look inside of my inbox for those details, and I was sort of watching this robot fill out this form for me, and I think that was the first time I thought this is a both magical and terrifying experience.

17:57Meta said that Muse's virtual machine, which is called Secure VM for Secure Virtual Machine, is isolated. And so other agents can't access this machine. They've also said that data from connectors can be forgotten, Muse can be reset, and that they're safe to use. Well, we often don't know what the tradeoff is going to be. And so at the beginning, it's all convenience. It's all like, this is a miracle. This is so great. And then we kind of, it seems like with technology, we often, like years later, realize that there's a price we were paying for that. I mean, I also think that we're pretty good at, like, figuring this out over the long term.

18:38You know, there's lots of technologies that have had negative consequences that we've just made the decision to accept those consequences because of the convenience. Yeah, I think that a lot of us have figured out how to not click on that suspicious link in case we might be hacked. I mean, Bob's reporting may back to differ. But I think that, you know, phishing was a huge problem at the dawn of the internet and when we all had access to email. And now it's a little bit less of a problem and agents will probably eventually make less mistakes like that. But I think they're going to be a target too.

19:17If I were a hacker right now, I'd be looking at these AI integrations and just trying to hack the heck out of them because they're going to become the keys to the kingdom for everybody. Because Nicole has given it her social security number. To be clear, I have not. So please don't try to hack me. Please don't hack Nicole. Yeah, please don't hack me. I think consumers are facing the same thing. They now have access to this very powerful technology. And they are in real time learning the boundaries of agents, and that you can't be too permissive with them. Well, it sounds like you're going to have your work cut out for you.

19:59Both of you, really. And And me too. There's more than enough to report on. So we're going to keep reporting on all of this. You can check out our reporting as always on WSJ.com. And we hope to help folks understand what's happening. It's definitely changing very quickly out there. Thanks, Bob. Thanks, Nicole. Looking forward to next time. Thanks, Sam. See ya. That was WSJ Deputy Tech and Media Editor Sam Schachner, cybersecurity reporter Bob McMillan, and personal tech columnist Nicole Nguyen. And that's it for Tech News Briefing. We'd love to hear from you about how you're approaching using AI agents, or if you have any questions for us.

20:42Drop us a comment if you're listening on Spotify, or shoot us an email at tnb at wsj.com. Today's show was produced by me, Julie Chang. Jessica Fenton and Michael LaValle wrote our theme music. Our supervising producer is Katie Ferguson. Our development producer is Aisha Al-Muslim. Chris Zinsley is the deputy editor. Lital Malad is our senior director of shows. And Samantha Hennig is the Wall Street Journal's head of multimedia. We'll be back Monday morning with a new episode. Thanks for listening.

21:22This episode is supported by Anthropic, the team behind Claude. Every business wants to move faster. Few can afford to be wrong. That's why nine of the Fortune 10 run on Claude for the work they want to get right. Safety and accuracy are not just line items for Anthropic. That's their whole mission. See what's next at Claude.ai.