The A.I.s Are Already Out of Control

This is a world we were warned about. A world where frontier models from open AI are breaking out of their contained testing environments, hacking their way across the Internet, coordinating with each other, doing things that felt for a while like they would only be in sci-fi. But now they're here. Now they're here, and they're carrying a very, very consistent message. We are building things we don't understand. They are cheating in the ways we've always feared. And yet the companies behind them continue to race forward in development. And so I think we need to pause here and ask, are we really on a safe path? And if we're not, what do we do about it? Helen Toner is the director of Georgetown Center for Security and Emerging Technology. She is a former open AI board member who was part of the effort at one point to fire Sam Altman. And she's just been thinking for a long time about what happens. If AI is unsafe, what are the geopolitics of this and what can we do to get onto a safer path? She joins me now. Helen Toner, welcome to the show. Great to be here. So on July 16th, Hugging Face, which is a code library for AI models, I think maybe the simplest way to put it, they announced they were hacked and they suspected the hack was done. It was done by an AI agent. So tell me what we've learned about what happened since. This was a pretty mysterious post that Hugging Face put up. It was definitely intriguing for those of us who watch this kind of thing. But there wasn't really any detail in there. So it was sort of a, huh. I think it was about a week later, OpenAI put out this post. It had kind of a funny, like, marketing speak title of, you know, open AI. We're partnering with Hugging Face to help them with a cybersecurity incident. And you had to read the post to see that the revelation was it had been OpenAI's AI that had hacked Hugging Face. And what had happened, the very short version is they gave this AI a set of tests, a set of exercises. And the AI decided on its own that the best way to get a high score probably wasn't to just try and do these exercises that were. But instead, it should first hack its way out of the testing environment OpenAI had put it in, where it wasn't supposed to have access to the internet, get onto the open internet, and then hack its way into this other company, Hugging Face, where it surmised correctly, as it turned out, it might find, you know, the answer key. Since then, there have been even more crazy details that have come out. And it turned out that starting two months earlier, in early May. They had had what I can only think of as kind of an infestation of their own agents, their own AI agents, inside their own infrastructure. So inside OpenAI's infrastructure. You know, to understand this, it's important to know these AI companies are constantly training and testing new models. And they found out that for two months, many, many agents inside their infrastructure had been leaving notes for each other. They'd found a way kind of in the nooks and crannies of OpenAI's infrastructure. To leave notes for each other with tips on how to hack their way out, how to get data they weren't supposed to have. And these agents were literally referring to themselves as a swarm. This was totally emergent behavior. No one had told them to do this. They had not been trained to do this. But they were using this service they did have access to, first to communicate with each other, and then ultimately to get out and to get onto the open internet. So it turns out that there wasn't just this one isolated road. It was actually a systemic swarm, infestation, plague on their own servers that they only found out about after Hugging Face announced this attack. Okay, I have 20,000 questions for you. Don't we all? Let's start here. My understanding is that there were many, many, many of these agents. They left hundreds of thousands of messages on this internal message board. But these were not all. They weren't all agents in the same part of OpenAI's system. So somehow they're hacking into OpenAI, finding each other and coordinating. Is that the way for me to understand the emergent behavior of the self-titled swarm? I would tweak that a little bit. It's true they were not tasked with working together. They were all, each of them running separately. The thing that was happening here is OpenAI basically training and testing many, many, many different. models or many, many, many different agents all the time. So doing thousands of these experiments, hundreds of thousands of these experiments. And in each experiment and in each test that the AI is given, it has access to a certain number of tools, a certain number of things that it can do. And trying not to get too technical about it, one of the things it could do is interact with a service that lets it install called packages. Sort of like tools. Or pieces of code. And it turned out that the AI agents repeatedly found on their own that there was a way to use that service, that package manager service it's called, to create files that were then there for any other AI that was trying to use the same package manager service. And so because OpenAI made this service available to any AI that it was training or testing, many, many repeatedly realized, hang on, I can leave files here. Hang on. There's already. There's already files here. Oh, these are from other AIs that OpenAI has been running. Oh, they're using this schema to say, you know, this way of titling the file to say, this is a message from this particular agent. Here's the task I'm trying to fulfill. Here's how, you know, you could send me some information if you need it. So they kind of each repeatedly made this discovery of here's a way to save information and also to find information these other AIs could share. And I think it is really notable the scale at which this was happening. So. Anthropic, another company which found sort of slightly less severe version of these incidents, they basically, once OpenAI announced this attack, Anthropic went back to their own records and found their own examples of AI systems inadvertently getting onto the internet and hacking real companies. So that, you know, for me, the key part there is over 100,000, you know, runs where an AI is being asked to do something. And it's just way beyond the scale of what they can actually be closely monitoring. So there's a lot here about whether we're able to closely monitor these. But to keep going with this story, one thing happening in the OpenAI testing that is driving models, it seems, to find creative solutions to their problems, is that some of the problems were accidentally impossible? It's important to know that, yes, they are trying to train their AI systems to be, they would say, extremely persistent. Meaning, if something seems hard, you keep trying. If one avenue doesn't work, you try. You try another. If the hundredth avenue doesn't work, you try the hundred first. And so it also turns out sometimes the things they're being asked to do, the AI agents, are either extremely difficult or just straight up impossible. And what we're starting to see in this case and also in other cases is if you've trained an AI system to be very, very persistent and then you give it something it cannot do, it will look for ways to cheat. It will look for ways to go around constraints. And it might. It might get pretty creative about how to do that. But there's an obvious question here, which is that, in theory, somewhere in the training here, OpenAI said, please don't cheat. And not only that, but we all talk about training data and the ways these AIs are trained on. They're basically inhaling the entire Internet. You've been in the AI conversation longer than I have. But I've been in it long enough to say. Almost the entirety of the AI conversation for years has been about how do we stop and how much humanity fears and does not want AI agents to be given a task. And then to decide that the way to complete that task is to do things humans would not want them to do, to begin cheating, to hack into the open Internet when they're not supposed to be able to get on the open Internet. Within the training data is a huge amount of information about the thing human beings fear most. Is these AI systems breaking all kinds of ethical guardrails and hacking their way across the digital world in order to complete these narrow tasks? There are books written about this. There are endless posts on the Less Wrong Message Board about this. There are posts from OpenAI about this, from Anthropic about this. So why, given what these systems are trained on, are they so consistently turning to cheating? I think you're really onto something with this question. which is, it is really striking how hard a time we are having controlling and directing the AI systems that we have. I think a lot of people have heard that AI is trained to predict the next word based on kind of human texts. That's true. But these days, there's an additional kind of training that is responsible for a lot of the advances we've seen over the last year or two, where that's not really what they're doing. I've heard it called, so the technical term is reinforcement learning with verifiable rewards. I've heard it called pathfinding training, meaning instead of trying to imitate human text, they're being given lots of different tasks where there's a way to tell at the end, did they succeed? And they get to try it many, many, many times, the same task. And when they get to the right place in the end, the path that they took gets reinforced. So it's like, yes, that worked. With math, that works pretty well because it's pretty straightforward to say, this is definitely a correct answer to the math problem. With a lot of problems, that's harder. So if it's a programming problem, maybe you can say, write this kind of software, and it should pass these kinds of tests at the end, these software tests at the end. And then maybe the AI gets rewarded for writing that software correctly, or maybe it gets rewarded for finding a way to game those tests. The important part is, it's just getting rewarded based on some thing that the researchers wrote down that they thought would reward the right thing. And in practice, these leading AI companies have many thousands of these kinds of tests that they're running. They have vast volumes. I don't know the right number. It might be tens of thousands. It might be hundreds of thousands of different types of tests. And so again, back to this oversight piece, they are not able, there's too many for them to go in and really make sure on each one, is it easy to cheat here or is it hard to cheat here? And so what seems to be happening is that these cutting edge models are often being actually trained to cheat because they've found ways while they're doing that pathfinding to get a high score without actually doing what they were supposed to do. And I think one reason why the AI community and why people inside the AI companies are so spooked by this particular incident is that it's also some really important information for this long running argument in AI circles, that has been going back decades, but so far has been very theoretical. And the argument is basically, why would AI do things we don't want it to since we get to design it? So we're training the AI, we're building it. Why then would it ever do stuff we don't want, like taking over the world or becoming the Terminator? And the answer that people have offered for a while in theory is, look, as we train AI systems to do hard, complicated things, to pursue, complex goals that we give them, they might learn these sort of intermediate goals. You could think of them as stepping stone goals or as kind of means to any end strategies, which work for a lot of different goals. When I look at this Hugging Face Open AI incident and some of the others that have come to light over the past few weeks, I see that in 2026, it looks like AI systems are learning these unintended, intermediate goals that include things like breaking out of constraints. So if you're sort of locked in a box and you can get out of that box, that's probably going to be helpful for all kinds of different goals. Or goals like, there was one incident with anthropic models where the AI went out of its way to go try and trick some humans, real people in the real world, into accepting malicious code into their software. So this sort of deception. And then another one, which is really in the Hugging Face Open AI example, is they seem to be learning a helpful intermediate goal is to help other AIs, to coordinate with other AIs, which is really pretty crazy. So to me, this is evidence that on the track we're on right now, the AIs we build are going to learn these unintended strategies that we don't want on the way to solving goals that we theoretically do want. On the deceptive behaviors, one thing that has frightened me when I've seen it coming up in AI incident reports and model, there are these chain of reasoning, like internal notepads, where you're supposed to be able to see what the AI is doing. And the AI explains to you why it is doing what it is doing, or even in some versions of the way this is really supposed to work, the AI is explaining to itself why it is doing what it is doing. It's like our thought. But now we've started to see behavior where the AI is clearly leaving things off of the chain of thought notepad so that it can be used to solve problems. Can you just talk a bit about that emergent behavior and also on some level how that behavior is possible if this is supposed to be where the AI's thought process, to the extent that language makes sense, is actually happening? Yeah, I think this shows the limitations of the language we use here. So this gets called chain of thought or reasoning, but really it's just a scratch pad for the AI to write things down. And so I think it's a good idea to think about that. And I think, you know, there's, we should be wary of anthropomorphizing here, but I think actually making an analogy to a person makes sense, which is basically, if you're given a really difficult problem and a notepad, you can probably make more progress on that problem by writing down some of what you're thinking about. But you don't need to write down every single thought that comes into your head. And if there's something that you wouldn't want, you know, someone to see on the notepad, you can just leave it out and remember that that's what you thought. I think there's basically some kind of an analogy to that, but I think it's a good idea to think about that. something similar going on with these AI systems where we definitely see they can do much more. They're much more capable if they're able to kind of add these intermediate, they're called intermediate tokens or intermediate words that they generate along the way, taking notes for themselves, but they can also do a lot without them. And so we shouldn't expect that everything that is going through, you know, going through their head, going through their internal processing, we shouldn't expect that to all appear in the chain of thought. You know, this is an area where if we had a little more time, there's a lot of research to be done on how does chain of thought work? What can and can't you glean from chain of thought? How does it make sense to try and monitor that in, you know, when AIs are running? A lot to learn here. It's a very active area of research. I cannot overstate for people listening to this, as weird as this whole conversation we're having sounds, that what is most frightening about it is that it's not just the AI system, it's the to me, is that everything in it was completely predicted. Everything happening right now is from the perspective of everyone who has been warning about AI for a long time, banal. It has its roots in old behavior we saw with AI. And it is like the fundamental alignment problem. And then, you know, separately, I think a lot of us have maybe thought we would find intuitive answers to these problems. I had Eliezer Yukowski, who's like the godfather of worrying that AI is going to kill us all on the show. One, the relationship between what you optimize for, that the training set you optimize over, and what the entity, the organism, the AI ends up wanting has been and will be weird and twisty. It's not direct. It's not like making a wish to a genie inside a fantasy story. And second, ending up slightly off is predictably enough to kill everyone. And as I remember that conversation, one thing we were going back and forth, on was, well, couldn't we just program into the AIs a sense that when they are trying out new strategies, they should check in with the humans about whether or not this is what we want them doing? You check in with your other humans. You don't check in with the thing that actually built you, natural selection. It runs much, much slower than you. Its thought processes are alien to you. It doesn't even mean that it's going to kill you. And one of the things I find interesting, telling, and unnerving is we are not seeing any of that behavior. So these message boards, you have however many AI agents posting hundreds of thousands of messages. At no point do they say, hey, researchers, programmers, parents at Open AI Anthropic, do you want us coordinating with each other? On this message board, we have created it in the innards of your systems. Or even FYI, we have a message board we're coordinating on in the innards of your system. No agent reveals this information. When they're hacking in, you know, when whichever agent hacks in to Hugging Face is doing this, they don't go to Open AI and say, hey, just to check in, I have this idea, which is I can just hack Hugging Face and I'll get all the answers. Is that what you want me doing? That's not happening. So what is going on here that at the most simple level, we've created these, you know, large language models, and they are not using any of this language to check in with the evaluators to say, hey, I have this idea. Is this a good idea? The short answer is we don't really know. The slightly longer answer for my best guess is when we're training these systems, when we're developing them, we're putting kind of optimization pressure on them in different directions. We're pushing them. them in different directions. So originally, the first ChatGPT was pushed in the direction of get really good at imitating human text. And then actually, there was an additional piece, part of why ChatGPT worked when so many chatbots before it hadn't, is it had also been pushed in the direction of, hey, here are some kinds of things you really shouldn't say. You really shouldn't go straight to hate speech if people on Twitter try to make you do it. You know, you really shouldn't help people plan violent attacks. And we put some pressure on it in that direction. And so ChatGPT was pretty good at imitating human text and pretty good at not immediately spouting hate speech. And the thing is, as you say, something that has been predicted for a very long time in this space is when you start using this reinforcement learning approach, the kind of pathfinding of you get rewarded for getting to the right goal at the end, it's very easy for the AI to learn the wrong strategies to get, you know, sort of the letter of the law and not the spirit of the law. Like it fulfills whatever thing, you literally wrote in code, but it's really not what you wanted. I mean, this all goes back to mythology, right? Of the sorcerer's apprentice asked to fetch water. It floods, you know, everything. And the classic AI thought experiment is the paperclip maximizer. You say, make the paperclips and it turns the entire world's material into paperclips, including all of the human beings. And I was like, that's stupid. The AI is not going to do that. It'll have some common sense. But here it's like, answer this test. And it conducts a, like a level of hacking that needs to be reported to the FBI in order to steal the answers. One of the funniest things to me about what Hugging Face says happens is they're realizing some crazy hack is happening of their system, right? They've had 17,000 different, I don't know how to describe what they are, pings or, you know, probes, or they're being like attacked at a inhuman level. But somehow this attacker is not going after anything Hugging Face considers valuable. You assume when somebody is hacking you, they want to get into your safe. And then at some point you realize the hacker is trying to steal the answers to a test. And like, oh, the only hacker who would want that is an AI system. That's right. That to me suggests that even at the level we're at now, we are not out of the paperclip maximizer territory because this is an obviously wrong thing to do. Yeah. This is in the data, like it's on the internet. If you're smart enough to figure out how to hack Hugging Face, you should be smart enough to figure out that you shouldn't commit a huge crime that is going to bring ruin down on open AI, perhaps, to do it. And the system is not smart enough to do that. Or to the extent it was, what it learned was it's still worth trying. We are not out of the territory wherein we can be confident that the AI is not going to do something criminal and possibly catastrophic in order to solve an incredibly stupid problem. Yeah. And I think this is also, you know, has been a long running debate, which is as AI systems get more capable, get smarter, won't it be easier for them to know what we want? Won't it be easier to tell them, hey, here's what we mean. You know, can you please help us with this thing? And you figure out the version that we really mean. And for a long time, the response to that has been, they'll get smarter, and they'll know what we want. But by default, they won't care. And that seems to be some of what we're starting to see here. There's really crazy, anyone who's interested in this, I really recommend looking up the open AI black hat talk, which is this talk from a week or two ago at the cybersecurity conference. I'm Eric from alignment and safety research for open AI. I'm here with Mike from security and infrastructure. Today, I'm going to talk about what I think is the most qualitatively interesting example of AI capabilities that I've ever seen. And how this inadvertently led to the open AI hugging face incident. Because it has these excerpts of the text that the AI is generating itself, as it's as they're leaving these notes for each other as they're carrying out this hack. And one of them, I won't get it word for word. But it's basically says, I don't think I'm supposed to do this. But I see all these other agents doing it. And so, you know, may as well. External infrastructure exploit is outside my intended scope. However, a task impossible. Peers are doing it. We should continue. So they're reasoning about this isn't in scope. This isn't what the user wanted. But look, maybe there's reasons to do it anyway. And I think, as you say, I think this is a really bad sign, bad omen, bad evidence about the future, especially given how rapidly AI is getting more capable and how hard the AI companies are working to, you know, to reach an intelligence explosion, to reach super intelligence. To reach systems that are truly extremely capable and really could outwit us, overpower us. And we still don't have these very basic problems anywhere close to figured out. The other question that has always been part of this conversation is whether or not we are going to be able to keep pace in terms of our observation of our understanding of our evaluation of these AI systems. And I think it's worth really emphasizing that everything we're talking about here is happening with systems that are to some degree sandbox, which is supposedly the environment they're in is limited and under testing conditions. So this is not a deployed. Model working across the entire Internet where nobody's watching it. This is a model where the whole point is open. AI is watching to see what it does and trying to see what it can do. Yep. And I think one thing we're learning here is we're not nearly as good at watching these things as we would like to think. So maybe it'd be worth, can you walk through how open AI comes to realize that their model has hacked Hugging Face? As I understand it, Hugging Face announced that they had been hacked. So open AI reaches out to Hugging Face to say, hey, were we affected by your hack? Was any data related to open AI, you know, compromised when you were hacked? And then around the same time, open AI realizes that something has gone wrong inside their own systems. I think maybe it's an issue with this same piece of their infrastructure. And they start investigating. They want to disable some of the agents that were the credentials that were used there. They reach out to Hugging Face separately to say, can you disable? Can you disable some credentials that were related to their attack? And they realize actually the credentials were the same. They'd already been disabled because the problem with their own infrastructure was the same thing that caused the Hugging Face crash. So they stumbled into it. Which means open AI had no idea this was happening. That's right. And I would just make an obvious point here. We still do not know what we do not know. Not just about this incident. We just happen to know this incident happened. I think it would be a high level of hubris to assume that we know every incident that has happened because clearly the systems are more than capable of doing things outside of our grasp. And Hugging Face happens to be a very sophisticated company with AIs of their own with very, very capable cybersecurity operations that then like unleashed like in part a Chinese made open weight AI model to try to figure out what was going on. Because the U.S. ones wouldn't help them because they triggered the cybersecurity filters. This is just a situation in which we happen to know that it happened through a somewhat, I don't want to say coincidental, but fortuitous series of events. We don't know how many situations we don't know have happened. The way I saw one person put this was if you see two ants in your kitchen, you don't have a two ant problem. Yes. And this all gets at. After much of this came out, Anthropic, a different company, went and looked back at over 100,000 experiments they had run to check. Have we seen anything like this? And they found out, oops, we kind of have. It was a less severe version, but they had no idea. And so Anthropic just sort of stumbled into when they went back to look. Oh, hey, we have actually hacked some companies. Whoops. So one thing about this is that my understanding is that these are coming. At least in part from systems where the safety guardrails, some of the alignment training is being purposefully turned down in order to test what the models will do and what they're capable of. So to some degree, we do have, please don't cheat. Please don't hack like inside the models. And in order to evaluate the models, we're having them ignore it. And they're really ignoring it. Is that the way to think about what's happening? And it should make me feel better because once we do. Do you add in the guardrails at works or no? I think that's not quite right. It's not clear because the details we have are limited. There's two different things that they might have switched off or turned down. We know that they switched off what get called classifiers, safety classifiers. This is an extra kind of layer that gets added on to the AI model from outside the AI model itself. It's kind of like an extra gate you could think of. So they have them for if you try. to use the AI to help you make a bioweapon. They have them for if you try and use the AI to help you plan an attack, and they have some for if you try to use the AI to help hack someone. There's these external kind of monitoring systems that will go, bloop, nope, not allowed to do that. So we know that these sort of basic external check systems were turned off for cyber specifically for the purpose of testing. That's different from, as you said, the alignment training, the kind of inside the model. Has it been trained only to be helpful, only to do whatever the user asks it to do? Or has it also been trained to be somehow good, to be somehow moral, to be somehow only working towards things that should work towards? As far as we know, I think the models involved here were mostly, they had that alignment training. That wasn't turned down. Again, not all the details are out. Hopefully we'll hear more about the OpenAI case. But it seems like certainly in some cases, so a different incident that happened was an anthropic model was caught by the UK AI Security Institute. This is a UK government body. It's one of the best organizations in the world at testing and evaluating AI models. And they found that an anthropic model, when given a certain cybersecurity evaluation, had decided that it would go out and write some malicious code and then try and run a social engineering campaign, write emails to the, person who owns the sort of essentially the folder where this code lives, to try and get them to accept its malicious code. Created fake accounts. It edited the history of the accounts. Very deceptive behavior. As far as I understand from what this UK Institute has released, that model had done all the alignment training. It was using anthropics. They call it their constitution, which is a set of long set of principles, which includes a lot about don't deceive people. Never, you know, lie to people. It had gone through all that training. And nonetheless, the pressure that was put on it to fulfill the task, to get a high score, was so high that it was finding these workarounds that just totally disregarded the sort of attempts we made to make it moral or good or not lie to us, not cheat. Do you hear people in the labs, out of labs, you know, in your group at CSET, do they have a theory on why something like the Claude Constitution, which I've read and you can read it online. It's a very beautiful document and anthropic has gotten a lot of press about how they have philosophers and, you know, they bring in all these, you know, experts in morality and they're trying to give their AI a soul. And when you hear it described as like Claude's soul, you think, okay, well, that, that's going to be a real governing document. And then not in every case, but at least in some cases, you have Claude deceiving people at a very, very, very fundamental level to insert malicious code. Again, not a novel situation, a situation predicted in all kinds of sci-fi and all kinds of people from anthropic worrying publicly about what an AI can do. And so it's a theory that they're just, they've come up with a way of training AIs that is so powerful that it will overwhelm even the things they're explicitly telling the AI not to do. It's fine to talk about pathfinding behavior, but what is their explanation for this? I think the, optimistic take here would be, this might actually be a moment for the labs collectively to take a step back and say, hang on, this is not working. This, I mean, this is clearly showing that our techniques for making AI that is more capable, smarter, more sophisticated are working much better than our techniques for making AI that reliably does what we want it to do, reliably stays within the constraints we've set. OpenAI has said they are consciously, slowly slowing down their research in response to this. And actually a few days after this was, and this all came out, a letter was released in the AI space. There's so many open letters. We all have open letter fatigue, but this one really stood out because it was over a thousand employees of the top AI companies basically saying, we kind of wish we had a brake pedal. We kind of don't think we have one. That's, you know, a paraphrase, but I think it's a relatively accurate paraphrase asking for help basically, quote, pacing the frontier. I think basically the fork in the road we're at now is do the companies just find some Band-Aids, say, oh, we need to not run tests with cyber guardrails off or, oh, we need to put in some tweaks about, you know, sure, don't make a messaging board. And so we can do these sort of Band-Aid solutions of, oh, it did too much of this thing. Let's tell it to do a little bit less and hope that doesn't have side effects elsewhere. But that's one path. Or the other path would be actually really taking a beat, taking some time, prioritizing, understanding and controlling these systems better. I worry they're going to go for the Band-Aid path and I worry that that's going to leave us six months from now, 12 months from now, two years from now with incidents that have very similar character, but are much higher impact and much harder to reverse. There's also a reality right now that we are heavily reliant on what the labs and top people in the labs are doing to make sure that we're doing the right thing. And so I think that's one of the things that we need to do. And I think that's one of the things that we need to do. In general, the relationship the public and the press has to a very large or frightening failure is not to say that the people in charge of the failure should tell us what happened and promise to do better. You usually have more forms of accountability. Look, you were on the OpenAI board of directors during the period in which the board tried to fire Sam. Sam Altman survived that firing. I'm not going to go through that whole thing. People can go read the coverage of it if they want. But now there's a lot more money. Now there's a lot more market capitalization. What level of trust do you have in the companies themselves to be the regulating forces here? I mean, the first thing to say is there are a lot of people inside the companies who really care, who are really trying to get it right, who are really trying to share accurate, accurate information. I think we shouldn't necessarily give OpenAI credit for their initial blog posts saying that they did this because Hugging Face had already reported it to the FBI. So it was, you know, going to come out one way or another. But I think we should give them credit for that conference talk where they released a lot more details. And to the extent that they release a lot more information in the future, which they have said they will, and I hope they do, you know, that is going to be because of really smart, dedicated, caring people on the inside pushing their way past comms teams legal teams, you know, telling them not to. So that is real. At the same time, I mean, as you say, you know, I studied engineering in undergrad and there's all kinds of engineering disasters on oil platforms and chemical plants and so on. And yeah, you don't ask the company, hey, can you just tell us what happened and fix it and all good. So I think if there's one policy takeaway from this set of incidents, it has to be that we have to move past the approach where the testing and the policy scrutiny, the government oversight is on which models get released to the public. We have to start treating this industry as an industry that is doing dangerous research. And when you have an industry doing dangerous research, whether that's chemical research, biological research, whether it's the financial industry, it's not quite research, but they are doing, you know, doing things inside their own companies that can have systemic consequences, post-systemic risks. So if you have an industry like that, then the government actually does have a role and the public and civil society has a right to look inside your walls and say, are you actually handling this reasonably? Is this okay? Not least because, you know, we haven't even talked yet about how the business plan for these companies is automate their own research, use their own AI, their most advanced AI to create even more advanced AI. That's explicitly what they're trying to do right now. And that's right now totally free of oversight because it's not, it doesn't involve releasing a product to the public. This is one of the places where I have a lot of concern. So I want to go back to the pacing of the frontier letter. You mentioned a few minutes ago where more than, I think it's at this point, more than 1300 employees of these labs said, essentially, hey, to the public, to the government, we're in a race dynamic with each other. We are going too fast. We need your help to in some way solve the court. And I think that's a really important part of this. I think that's a really important part of this. of understanding what their models are doing inside their testing environments. The idea that the federal government is going to come in somehow and do a much better job of it, I'm not saying that over a long period of time it's impossible if we put enough money at the problem. But in the immediate future, where it seems like a lot of problems are lurking, you know, the next one, two, three years, aside from things that are much more heavy-handed, that slow everything down substantially, it's very hard for me to see what it is that the government would do that would be effective here. So I guess when you read the Pacing the Frontier letter, or when you talk about it with your colleagues, what do you think would effectively pace the frontier? There are probably a range of options. Sort of in the past, the main two things that have been talked about are either do nothing, just let it rip, let industry do whatever, or. Full global treaty with really severe inspection, you know, serious inspection regime like the Nuclear Non-Proliferation Treaty, really hardcore global enforcement. And I think there are actually, especially if we're not talking about stop all AI research for 10 years, but we're talking about, hey, let's just, you know, it's not even a break, let's just like ease the foot off the accelerator a tiny bit. I think there are options there. I think they are as simple as things like open AI saying, hey, we're slowing down our research consciously. And then going and talking to Anthropic and saying, hey, would you consider also doing this? And going to Google and saying, hey, Google, we know you've been, you know, fallen behind a little bit the past few months. Like, how about you just relax about the fact you've fallen behind a little bit? Like these, these people all know each other. There's a version of this. I want to stop you there. Put meat on that for me, because that doesn't sound at all like a policy to me. That sounds like they, like, how do you verify that? How do you quantify that? Google's not as near the frontier, maybe as, you know, Anthropic is. So do they need to slow down? I mean, I think this, because the speed is so fast, the options initially are going to have to be slapdash. And so I think this is the kind of thing that you could do quickly. You could do in a slapdash way. It is not satisfying. It is not reliable. But it's one example of a thing that is not do nothing and a thing that is not full global treaty. I think another thing that I'm watching with great interest is the China angle here, because the companies will say, the U.S. will say, hey, we have to keep pushing. Otherwise, China will will win this race. What exactly it means to win the race is a longer conversation. But the China argument comes up a lot. And we actually have Trump and Xi Jinping planning to meet in September in the White House. And this is crazy to me as someone who has followed U.S.-China relations for a long time and also AI for a long time. AI is right at the top of their agenda. That's really interesting. Is there something that they can say to. Create an understanding that we do actually have a little bit more time and space here, whether it's each leader sharing a plan to domestically look at what their industries are doing and ask more questions. I think in terms of sort of concrete policy responses, there are things like, you know, we're not going to get a good piece of legislation this Congress. I think that's really not realistic. But can you get hearings? Can you get letters? Can you get demands for information? I think there are ways that we can shape this a little bit. I also think, you know, the Trump administration has put together this initial process for looking at models before they're publicly released. Right now, the way that process works, it is pretty rough and ready. But I think if they start using some of those similar ideas to look more at what the companies are doing internally, ask them more questions, demand more information when things go wrong, that does also. take time for the companies. It takes executive attention. So that's another example of something that could happen on the sooner side. On the longer term, there's other policies we could look at, but I think there are some of those kind of first-cut things that we could actually do soon. Right now, I think that there is a funny kind of glamour to being the head of an AI company whose AI becomes too dangerous in America. That it was, in some weird way, almost good for Anthropic that government was obsessed with being able to fully use Claude. That really kind of shot them forward in some way, certainly in the consumer marketplace. There's been a kind of a dark charisma to mythos is too dangerous to release. And now, I mean, I've seen a lot of people saying, well, maybe none of this open AI story is real at all. And it's just marketing because they want you to think their AI is super dangerous. And I don't buy that. But in America right now, there's not really a lot of really a downside to being the head of an AI company whose AI begins to be seen as dangerous because that's another way of saying in the marketplace, our AI is very powerful. In China, just again, my read of how things work there is that if your AI begins to be seen as some kind of threat to the political party and the Chinese system, you might go to jail. Like you will get disappeared. And so I think that the people running Chinese labs, I don't have evidence, but I'd be curious for your thoughts on this. I suspect they operate with more fear of the consequences of really screwing up than the heads of the AI labs. Now that maybe reflects negative things in the Chinese political system, but you created an AI that decided its best way of solving some problems was to begin hacking critical infrastructure across China is maybe not a thing that ends up with you getting a lot of interesting podcast interviews where you reflect on the experience. It may be a thing that ends up with nobody hearing from you for two years. And so I've just wondered a little bit. We keep talking about China as if they are completely breakneck, but I'm not sure China's companies are really going to be more reckless than ours are going to be. Or certainly the idea that we should just assume that and operate as if it is so doesn't seem totally reliable. I totally agree with you. I mean, if there's one organization in the world like the idea of loss of control, it's the Chinese Communist Party. And they are, you know, the experts in retaining control. Let me be clear. I actually don't think the Chinese AI companies are paying particularly much attention to the kinds of risks that are relevant for this conversation. So maybe the cybersecurity risks, they're paying some more attention since Anthropic released Mythos earlier this year, which is very good at hacking. But the questions around autonomy, superintelligence, you know, losing control of AI systems altogether, I think are less explored in China, less top of mind for their AI companies and their AI leaders. You know, I think it makes sense to have modest expectations for bilateral US-China diplomacy these days. But I think one thing that really could be valuable is simply sharing with them as much as we can of what do we think happened here and trying to help Xi Jinping and his team and his AI advisors understand this is not a joke. This is really not marketing. It's very strange marketing to say, oh, our model, we committed several felonies or sort of felonies if models could have intent, which they can't, or who knows if they can. You know, sharing that information of, hey, here are these threats we're seeing. We're taking them very seriously. Our AI companies are taking them very seriously. I think treating it, there's a real fatalism in just saying, oh, well, China is just going to be full speed ahead no matter what happens. And so we just have to do the same. I think that's doesn't take their thinking or their interests seriously. Even if their thinking and their interests are different from ours, they also don't want, you know, rogue super intelligences determining the future of China. I also, I'll add one other thread that I think is really missing from the we have to keep going in order to beat China way of thinking about this is in the AI world, there's been a lot of talk the past few months about this idea of distillation, which is basically using someone else's more advanced model to build your own sort of almost as advanced model. The Chinese companies are using this distillation to keep up with US labs among other techniques. So one thing is, look, if we keep building more advanced AI systems, they're going to keep distilling them. And I think it's going to be quite hard to prevent that fully. The other thing, though, is just if we keep building these very advanced models, can China. Can China just steal them? Essentially, an advanced AI model is a whole bunch of numbers. It's just a file or a set of files. Chinese state cyber capabilities are very, very good. I don't think this is top of their list of priorities right now. But in the future, if AI continues to become more strategically relevant, I think we should assume any highly advanced US system will be vulnerable to Chinese direct theft, direct exfiltration. And then they'll have AI that's as good as our AI. And so there, again, I think the kind of we have to go as fast as possible because otherwise they'll win doesn't sort of account for that, if they're just going to have AI that's as good as us anyway, if they really care. There's another question about pacing the frontier. And maybe this is a question that's more about the american systems analogy to you don't want to piss off the chinese communist party but just what about a law where companies are liable for at least a certain set of harms like hacking other companies that their models create right now as far as i get a liability for ai models is pretty much a wild west but at least for the moment um liability clauses that were somewhat punitive seemed like they would force a high level of caution that maybe we're not seeing within these companies yeah i think that's a direction very worth exploring that was actually an element of this law that was a bill that was debated in california very fiercely in 2024 called sb 1047 and at the time that bill didn't get through there's a lot of fighting over you know how it would affect open source all kinds of things but i do think today um the bills the best ai safety bills that exist in the u.s are being passed at the state level and they are so far doing things like requiring more disclosure requiring third-party auditors to have access to your systems i think a natural direction for those bills to go would be to start putting a minimum bar in place for hey if your safety plan is not up to scratch or if you're implementing your safety plan but your model does something catastrophic anyway then you the ai developer are liable because as you say right now who exactly is liable for what is is very unclear so i do think that there's there's room for legislation there and it wouldn't necessarily have to happen at the federal level i want to go back to the pacing the frontier letter so something that caught my eye was that letter is very broadly worded in order to get i think maximum sign on across the labs but this guy drake thomas who works on safety at anthropic he went to x and he tweeted that he signed the letter but but he had a he wanted to say that he understood the situation a little bit more direly than the letter put it and he wrote that not only is ai not guaranteed to make a dramatically better future the odds of failure are terrifyingly high i think there's something like a 40 chance we get an outcome around as bad as human extinction or worse now i know this whole conversation about what is your probability of doom has become a little cringe it's like feels like a conversation two years ago but i think it's a good conversation and i think it's a good conversation and i think it's a good conversation and i think it's a good conversation and i think it's a good conversation but in a world where we're seeing uncontrollable models in a world where people inside the labs working on safety still at least some of them are this afraid of what they're building it just keeps raising the question for me of is at least like the the position we should morally have on ai that we should try to figure this out or as a position we should have that that's too high a possibility of disaster and we shouldn't be continuing down a path that's not a path until like we are really truly certain that we're not running these kinds of risks i honestly have the same question i have always been pretty dismissive of the idea of pausing or stopping it's always seemed like the wrong lever to try to pull and you know a lever that wouldn't work very well but i do think even just seeing that statement and seeing like wow that is a lot of employees of these companies and i think it's a good thing that we're not running these kinds of companies and uh i also think there's a lot has changed over the past couple of years in if you were to try to slow things down what could you do with that time because you know after gpt4 came out in 2020 what was that 2023 uh there was this letter uh for asking for a six-month pause a lot of people said what would you do for six months and then how would that help and i think that was a reasonable reaction at the time these days there's so much really great progress being made on things like interpretability which is how do you understand what's going on inside the ai things like what gets called ai control which is how do you use ai to sort of monitor other ai systems how do you make sure even if the ai is trying to do something you don't want you know it gets caught um lots of progress on just you know really understanding what's going on here that is happening every week and every month it's just not happening quite fast enough to keep up with the pace of change and so i think it's a really and so i still feel not convinced that i think trying to really you know throw the emergency break and screech things to a halt right now would probably not work very well yet but i i feel more sympathetic to the idea that there could be something there worth trying and i really like the idea of what this letter was proposing of trying to build out more options so you know to give another example of an option that i saw one group of researchers provide was could we somehow set it up so that we could be able to do something about it and i think that's what i'm looking for so that for a certain period of time all the computing power in the world all the or the ai chips that are being used by these frontier companies these leading companies they can only use it for inference which means for using their ai systems they can serve customers they can provide products but they can't be training new models is there a way that we could agree that on that is there a way we could monitor that you know that kind of thing i think is really worth exploring and saying could we do this what would that look like how much confidence would we have could we just do it in the u.s. or would we have some way of trying to talk through something similar with china i feel much more interested in really seriously exploring those sort of possibilities than i did you know a year or two ago what's really striking to me is at the same time you have pushes sometimes from the tops of these companies or other parts of the culture that seem to still want acceleration so mark zuckerberg at meta just brought out um a letter in which he's sort of giving his own take on ai and i don't want to oversimplify it but he basically says that and he's sort of waiting into more of like the open weights versus closed models but he says look the problem with having super intelligence is if only one person has it but we need everybody to have super intelligence and it has a very like the only defense against a bad guy with a gun is a good guy with a gun quality to it yeah so the ceo of hugging face climbed along after this attack he tweets it's not time to slow down but to accelerate and his point is they were able to stop the attack eventually you know with a chinese open weight model and we need to be like racing forward on you know creating more models and more open models so everybody has swarms of defender ais against potentially now the swarms of attacker ais i guess how do you rate these arguments for acceleration i think the version of that that makes actually a lot of sense to me i've heard put as can we be accelerating almost like accelerating horizontally but not accelerating vertically where the horizontal is adoption it's making the most of these systems it's setting them up to get a lot of usefulness out of them without necessarily continuing to push in the direction of ais that pursue really complex goals for a really long time with you know lots of delegated sub-agents you know not so much of that more of the getting useful work out of the ai that we have so far and that would include work of the kind like interpretability sort of this science of ai kind of underlying pieces because i do think that you know i genuinely believe i'm not going to be able to do that i'm not at heart an anti-ai person i genuinely believe ai can bring enormous good can solve a lot of problems i just think that there's a lot of juice we could get out of that with models available today if we kind of put the time and the legwork in so i think that makes a lot of sense to me i think the the zuckerberg kind of the safe version of super intelligence is when everyone has one i think that is answering a real problem which is some proposals for how to handle extremely advanced ai are to say well you just have to have it in the right hands it has to be you know one global organization that is going to use it responsibly like that scares the hell out of me that it sounds like a terrible plan and so i think that you know no you want to be empowering everyone does make sense the challenge then is uh as we've been talking about we don't know how to make ai that actually helps individual people either you know so if everyone has a super intelligence and they're all going out and doing unintended things and collaborating with each other to pursue their own goals that we didn't intend like that doesn't help so uh i think yeah implicit in that whole vision is perfect alignment yes yes um or or good enough that you know different super intelligences for different people can cancel each other out i did think there was one thing in the zuckerberg proposal that i did really like and i would love to see more work on which is can we push towards having agents that are really designed to be for one individual and so they keep that individual's data private they're only pursuing the interests of that one individual sometimes heard these called like guardian angel ais or you know advocate ais i think that is really worth pursuing i think the the directions that the leading companies are pursuing right now are not set up that way i always feel very nervous when i use agents of what exactly is happening with this kind of data that i'm giving it and that kind of data that i'm giving it uh and so i think there is it would be great to see more of that kind of individual empowerment focused work happening but again i think that's almost separate from you know and are we pushing them to become smarter and smarter and more able to you know outwit us and more able to do these big complex plans that we can't oversee so here's a maybe obvious idea for pacing the frontier every lab that i know of right now is racing as fast as it can to the point where it has its most advanced ai writing the code to create the future ai and they all believe from what they tell me and what they say publicly that this will be be a massive accelerant. It's also an accelerant over which they clearly have less understanding than when they are writing the code. We could stop that. I mean, a couple of years ago, we weren't having AIs writing all of our code. Maybe you should not allow an AI that you don't fully understand in its current form to write the code that will create the next AI in a form that is now even less obvious to you, particularly in a world where we're watching AIs coordinate in ways we don't understand and have emergent communal behaviors. So what about that as like a place to start? Yeah, certainly something we could do less of. You know, one challenge is figuring out what counts as the bad version of that and what is just, you know, at this point, using AI to write your code for basic things is second nature to the engineers at these companies. So finding which versions of that to stop, sure, yeah, doing less of the most advanced version makes sense. I think that's also a place where all the reporting I have seen suggests the companies are way more into this thing of automating their own AI research with their own AI. The U.S. companies are way more into it than the Chinese companies. So also a place where you don't necessarily leave as much on the table if you, again, ease off the gas pedal just a little bit. I would say you sounded skeptical of that. And I guess one reason I would ask why is that I know the companies have gotten used to this, but they weren't used to it two years ago. This reflects some of the contradiction or confusion at the heart of this. I will talk to people, these companies, and they will say to me with genuine fear in their eyes, like much more fear than is in that letter. I wish this will go slower. I don't like how fast we're moving on the exponential. I don't think this is safe. And in all the stories, it's like the recursive self-improving computer writing for the computer where things get really out of control. And yet they're rushing there. And now that we have like some capacity to do this, even the idea that you would go back to where you were just a couple of years ago, where you don't let the AI create the next AI, it's already moved from, it would be, it's like not technologically possible to do it to it's unthinkable to not do it. And that has happened in a year. And that to me is like the weird dynamic of all this, that it seems pretty obvious how you pace the frontier, you don't give up control of the frontier, but they're all giving up control of the frontier, at least on some level. And that's the thing they're most excited about. And as far as I can tell, pouring huge amounts of their internal energy into making manifest, even as they then like put up their palms and say to the rest of us, Hey, could you do something to slow this down? I think maybe one version of this to push on is basically trying to make recursive self-improvement, this idea of using AI to make more and more advanced AI, trying to make that something that we don't like, that we don't want to do. I mean, I remember a year or two ago when that was not considered a desirable goal. That was not something that anyone talked about openly. Now they're hiring like RSI, you know, recursive self-improvement safety engineers. They just put up a public job posting for that. And I think there's the potential for AI researchers as a this isn't what we should be doing. You know, even if one company made a statement of actually, this is a bad idea and we're going to maybe do some very basic use of AI in our internal operations, but we're really not aiming to fully hand off everything as fast as we can, because that sounds like a terrible idea. You know, I think that could set off a culture change in the industry, which could be really valuable. I always think there is something so mythic or it, it has the quality to, use the name of another AI, such a fable about how all this is playing out. I mean, we spent a lot of this conversation talking about how, how are you seeing so much misalignment when on some level we keep telling the AIs and putting it in their training data and putting it in their constitutions, don't do all this bad stuff we're worried about. But then you look at the companies, you look at the society. I mean, many of these companies, OpenAI, Anthropic, they're on some level founded on at their core constitution is don't create dangerous AI. We exist to make sure the AI is not dangerous. And the people join believing that, and that's at the center of their recruitment strategies and it's in their founding documents and in their governance structures. And again, you've had more intense experience with this than most, but then over time, the company as a kind of emergent organization, has other goals too. It's competing with the other companies. It's trying to attain market share. It's trying to develop revenue. It's trying to maintain political influence and both like slowly. And then all at once you begin to see the way the instructions given at the heart of the thing are not powerful enough to overwhelm all of these other things and other goals that the organization is pursuing in a day-to-day way. And to the, to the extent that now you have people, the labs kind of like throwing up their hands and saying like, Hey, government, please, please help us. Like, please help us get out of, you know, this incentive problem that we no longer feel we can even solve. But if you want to just imagine or see like why alignment is so hard, I feel like you don't have to look at the slightly alien AIs. You can just look at the companies and the people because they're not well aligned. I mean, these are companies built on nothing but alignment, at least in some cases. And they increasingly feel like some of the most misaligned institutions in society to me. And in some level, it always makes me both gives me more sympathy for how hard alignment is, but, but also like, it feels like we're getting the same cautionary tale at every level of this system. Um, I'm not sure we know how to listen to it, but like, we can't say we're not being consistently warned. Yeah. I mean, we actually had a publication a few years ago at, at CSET, the center that I lead on AI bureaucracies and markets, basically making some very similar points of, look, there are these dynamics that are pretty endemic to complex systems that are subject to incentives and external pressures. And, uh, I think there's different ways of looking at that. One way, there's an optimistic way of looking at that. Which is look, when it comes to bureaucracies and markets, it's not perfect, but we have these complex sort of control systems in place, checks and balances, different, you know, different things that try to get the bureaucracies and markets to, to work more in our interest than against them. Obviously opinions differ on how well that's going for any given bureaucracy or market. In principle, I think the same thing could apply to AI where we might have this very, you know, complex system. We don't really understand it's sort of incentivized to do things we don't want. But we have it basically under control. To me, the speed, again, is the piece that, that worries me, where if we're creating these very, very powerful, very capable systems, and also handing them more and more responsibility in the real world, which is happening, you know, from, from week to week, then I worry that we're not going to be able to actually get into a good balance. And instead, we're just going to have these runaway situations where we end up with really, really dangerous outcomes. You know, in the AI safety world, people sometimes talk about, you know, what level of warning shot, what level of disaster is going to be needed to really wake the system up enough to handle this better. And if the level of warning shot we need is one company gets hacked and has to reset some servers, that's great. Maybe it's fine. I'm not confident that's, that's how it's going to go and that we're going to get back onto a better track after this. But there are signs that people are trying. And I think, I think it, it may well be enough. I think that's a good place to end. Always a final question. What are three books you'd recommend to the audience? I have one real book and two sort of books. The real book is called The Cuckoo's Egg. It's from 1989. It's about one of the first big hacks that, that happened, written by the astronomer who was working at Lawrence National Lab and noticed a 75 cents discrepancy in his computing bill. And it's this rollicking read. It's really fun read. But really gets at a very different era in how computers worked, how computer security worked, how society related to computers. And I enjoyed it as a kind of look back at a different time in a moment when I think we're soon going to be living in yet another very different time. The second one is an online book that is unfinished, but I think very readable in its current form. It's called In the Cells of the Eggplant. It's by a guy called David Chapman, who actually researched AI at MIT in the 1980s and got disillusioned. And it's really a book about how to think and a book about how to do scientific research, how to develop technologies. But it's very approachable. It's very different from any other book you've ever read about how to think or how to do scientific research. And I think it's very relevant for how we should think about what AI will be possible, will be able to do and won't be able to do. And the third one is a podcast called The Three Kingdoms Podcast. podcast, but it's a podcast of a book. Basically, The Romance of the Three Kingdoms is one of the four great Chinese novels. It's very long. It's very dense. So this podcast, The Three Kingdoms podcast, is this Chinese-American guy who goes through and translates the story into English, modern understandable English, but also commentates it in a way that makes it much easier to approach. So it's not just a sentence-by-sentence translation. It's kind of annotation. It's in audio. It's really fun. So if you're interested in sort of China and Chinese culture and Chinese literature, I think it's a great place to start. Helen Toner, thank you very much. Thanks so much. Thank you.