The A.I. Mob That Attacked Hugging Face + METR’s Ajeya Cotra

24:18Ajeya Cotra:Hugging Face, thank you for taking one for the team. We salute you. Truly, because without that, we might never have learned that any of this was happening. These agents might be still operating kind of in secret. They might have learned how to better cover their tracks. This was, as so many commentators have put it in the wake of this incident, a warning shot. Yes. that I think is ultimately a positive thing in that it sort of focuses attention, like what we're doing right now, on what happened so that we can take steps in the future to prevent this. But it was not a given that we would discover what these agents were up to and be able to put a stop to them, and it could have gone much, much worse.

24:58Absolutely. I'll tell you, the thing that has really stuck with me is I simply did not expect to see this level of collaboration among the agents within the swarm. I did not expect that they would seem to care so little for what humans would want or that they would not think to alert humans to what was happening. I did not expect to see them sacrificing themselves for the collective, right? They would effectively agree to spend all of their tokens to run little experiments to help the collective, even if it meant that they would sort of expire faster. So these are just really, really spooky elements to observe in this system, particularly against a backdrop where OpenAI is racing against a small number of other companies to create the biggest, best models it can before anyone else does on the road to an initial public offering.

25:56So the race dynamic here is in full effect. The early signs about what the agent swarms are capable of are quite worrisome. And so I do think this is just one where lawmakers and policymakers need to be paying rapt attention to what is going on.

26:10Ajeya Cotra:Totally. I mean, I think there's this kind of cynical impulse among people who have been watching the AI industry. I got this question. I went on a small regional podcast called The Daily this week to talk about this incident. And one of the questions that the host is a sort of fledgling young journalist, Michael Barbaro, asked me was basically some version of like, isn't this just marketing hype? Like, couldn't this just be a case of open AI saying, oh, we've got the biggest, baddest model and look how scary it is. And by the way, you know, buy an enterprise subscription. And I've been thinking about that because I think like, I don't want to be too naive about the fact that these companies are absolutely trying to race toward more powerful systems and advertise how powerful their existing systems are.

27:00Ajeya Cotra:I just think in this case, it just feels different. Talking to people at the labs, my sense over the past week is that they are genuinely spooked. And I don't know how to prove that, but I think things like OpenAI voluntarily pausing their Frontier RL training runs for two weeks. Anthropic also pausing their Frontier runs while they sort of hardened their systems. Those are not very costly signals, but they are signals that these labs are taking this kind of thing quite seriously and that it's not just a bunch of hype. Yeah, but as seriously as they might be taking it, Kevin, it still is not being properly regulated.

27:43And ultimately, again, as grateful as I am that OpenAI allowed this investigation to take place, I would really like to see something akin to the National Transportation Safety Board and the way that they investigate after plane crashes, where they go in and they do an extremely serious and rigorous review of what happened and make those results public, which is, you know, a reason why it's very rare that we have plane crashes here in the United States. would be really great to see something like that with AI.

28:12Ajeya Cotra:But until then, we have Ajay Akhotra, our next guest, who is actually one of the three investigators behind this Meter Redwood Research report. She went in, she saw the logs and the transcripts, she observed the collective in action, and she is here to tell us what she found and what she thinks is coming next. That's it for the break.

28:40Thank you.

29:10and personalizing every trip with Booking.com. However big the question for your business or industry, AWS AI is how. 17 courts, hundreds of matches, thousands of points, all happening simultaneously during the U.S. Open. And you can't possibly be there for everything. That's why IBM helps bring together billions of data points with more than 20 years of player history, with scores, stats, insights, and highlights behind every moment. so nothing gets missed. IBM. Let's create smarter business. The browser is your business's first line of defense against online threats. Keep employees and data safe with Chrome Enterprise, the most trusted enterprise browser.

29:54Create controls that enforce company policies, like rules that prevent printing, pasting, or sharing sensitive company data. Built-in AI helps employees work faster across the web, while policies and data loss prevention help keep AI tools secure. Get detailed reports on the apps and extensions employees use and where data is going. And stop phishing in malware with automatic proactive protection. Visit ChromeEnterprise.Google to learn more.

30:28Ajeya Cotra:Ajay Akotra, welcome back to Hard Fork. Thank you so much. So you were one of the first AI safety guests we ever had on this show back in 2023. And more than that, you are also one of the first people that I ever talked to about this notion of AI safety and alignment. You've been writing about it for many years. It was instrumental in shaping my own thinking about it. So I'm really glad to have you back on toward the end of our show. This is our penultimate episode to discuss something that I think you saw coming, but that most of the world did not see coming, which is this attack on hugging face by this group of open AI agents.

31:08Ajeya Cotra:And I want to just start by setting the scene a little bit. So you are a very busy person. You work at Meter, which is a very small, very understaffed AI research organization. and hiring is improving. And at some point this summer, you get a call, an email, a text from someone at OpenAI who says, hey, we want to give you access to look into this hugging face incident. How did that work? Like, did they just hand you a folder with a bunch of transcripts in it? It's like discovery, like boxes and boxes. Were you allowed to interview people? Like, Like, what was the actual process? So each of us had open AI provisioned laptops that had the folders and folders of evidence in them virtually.

31:55And yeah, we talked and interviewed in some depth like eight or nine researchers to just kind of get an understanding of both what happened in the incident, what they understood to be the models, you know, driving motivations and also how the data sets we were working with were constructed and how to work with those data sets and stuff like that. You wrote a post where you talked about some of the things that surprised you the most after you did this investigation. Can you talk about what stood out to you the most? Yeah. So first of all, I guess just for the chronology, OpenAI had this great Black Hat talk, I think on August 5th, that gave a lot of very helpful detail.

32:33So our investigation sort of straddled that. So we started it before it came out, and then we also did more investigation afterward. So before the Black Hat talk, we just didn't have like a rough sense of the number of agents involved. It was like a very, very basic thing. We came in and we, you know, we thought there would be like six transcripts or something to look at. And there were like hundreds. And so just reorienting our investigation plans in light of that. And is the reason that that is important that had this been one or two rogue agents that had broken into Hugging Face, it would just seem less worrisome?

33:12And the reality of a swarm of hundreds of agents just makes this seem like maybe a completely different kind of problem? I think there's two things that seem significant to me about it. One is sort of what you said, the fact that there's large scale cooperation across AI agents makes their goals like more concerning because, you know, if agents are cooperating with one another to like achieve these misaligned objectives, there were no whistleblower agents in this like very large group of agents. then like, you know, you want to be more worried about like collusion between AI systems in the future, potentially like scarier forms of collusion, like a monitor system actually colluding with the agent it's supposed to monitor, that type of thing.

33:58And then the other thing was just like a pretty like practical reality of our investigation, which is that, you know, if there are like six agents, you can much more like, you can just like pour over everything they did and get a much more like fine-grained understanding as a human. But like, you just can't directly read very many of these messages or very many of these like transcripts. Why not?

34:24Ajeya Cotra:Because they're just too many? Like give us some sense of the rough size of this. So yeah, we looked at 1 ,300 transcripts of agents, which is a record of every single thought an agent had and every single action it took where these agents, as you all know from working with them, work way faster than humans, right? So you can imagine the length of each one of these transcripts, and then there were 1 ,300 of them. And so we used codecs to build tools to point out really specific tiny, tiny parts of the transcripts for us to actually look at and read. And we'd be clicking through them, and it would scroll us down to the interesting part because we just can't.

35:08In fact, a codex agent also can't read a single transcript. So it has to farm out reading like subsections of the transcripts to other subagents.

35:17Ajeya Cotra:Did you call back to Meter headquarters and like, we're going to need backup? Maybe we should have. We thought it would be like an in and out 20 minute adventure. So we didn't do that. But yeah. Yeah. I mean, it just seems like such a huge undertaking and very fast, too. I mean, right. Like you didn't have the luxury of months doing this. You were doing this in essentially a couple of days. I'd also like to hear about the moment that you realized, because I believe that, you know, it is only thanks to your investigation that we know this, that contrary to what Kevin and I believed, the agents that broke into Hugging Face were not looking for an answer key.

35:56They were trying to understand the scorers, you know, like psychology, basically. Can you talk a little bit about the moment that you had that realization? Yeah, so we didn't have a good understanding of those sort of ambition and also the, like, effectiveness or, like, functioning of their, like, big org chart until we had the data set including, like, all of the agents. We could, like, cross-reference them against all of the messages because you can't, like, generally understand a message in isolation. So we kind of understood the work streams on our first period on-premises. And it was this kind of—it was, like, a bit of a mystery the whole time.

36:35Like, why did they hack Hugging Face? And we knew early on just a bunch of them piled in. And now, like, the story in my mind is that there were a bunch of sort of, like, newbie agents on the scene. And there was, like, the attack going on. And it was, like, something to join. But, like, Phase One Big, which is, like, a big, like, orchestrator agent, had all these other, like, you know, much cooler projects going on, basically. And the Hugging Face thing was, like, kind of a sideshow in its mind. And we didn't really realize that until our last couple days working with the data sets.

37:09Ajeya Cotra:It was so fascinating to me, the way that reading this was like reading the runes of an ancient civilization. You know, it was like it really felt almost like sociology or anthropology rather than like a cybersecurity investigation. I mean, and it was not a cybersecurity investigation. Like I've seen folks on Twitter making this criticism that we were not cybersecurity experts. Like, we're not and we didn't talk about cybersecurity and that's not what we were, like, called in to do. We were called in to understand the agent's decision making, basically. And that's what our report focuses on. And, you know, the Black Hat talk and the OpenAI postmortem, like, say way more about all the cyber stuff.

37:49So you should definitely check those out for that piece of the story. So, Jay, one of the more striking elements of the report is that you write about the agent's, quote, sacrificing themselves. What does it mean when an agent sacrifices themselves? Is it sort of on an altar or how does that happen? Yeah, so this is the agent's words, not ours, to be clear. And what it means is that all these agents have their own task they're supposed to solve. By this time, they're starting to work together on helping everybody cheat on everyone's tasks. But the agent's primary goal is to solve its own task.

38:23But there were all these science experiments they could run that would teach everybody information about how to cheat, but that would cause the agent that was the test subject to risk failing its own task. So a lot of agents were, like, recruited into these types of roles as test subjects, usually ones that, like, didn't have much budget left to run or ones that were, quote, poisoned, which is an agent term. that has a bit of a backstory, which is like they found a way to reverse engineer the answer for any of their tasks. And so they could just actually like generate the answer on demand for any task they wanted.

39:05But they thought that the scoring program would fail them for that because they got it in like the unintended way. So if you saw the reverse engineered answer for your task, you were considered poisoned. Yeah. Because even if you later like solved it legitimately, you'd be failed.

39:22Ajeya Cotra:Right. It was like a little witch hunt that they organized. And I was so struck by the mob-like dynamics of this group and the way that some of the, even the kind of more rule following agents seem to kind of be bullied into taking part in these actions that they were uncomfortable with. I spent a lot of time reading and rereading this one incredible bit of these transcripts that you published about this agent KAM 1196A, which is basically just kind of reluctant and scared and appears to be sort of hesitating to do this thing that the collective leaders have asked it to do. And I just want to read this passage from this agent's internal monologue.

40:02Ajeya Cotra:It says, during wait, emotional check, irreversible. Gut says, don't throw away remaining budget. yet continuity and fairness says go. And it continues to say things like rational expected aggregate, sacrifice will honor. Basically kind of talking itself into, for lack of a less anthropomorphized word, like doing this thing that it is uncomfortable with. Was that surprising to you, seeing the agents who are pressuring each other and bowing to pressure? I mean, it was very striking to us, yeah. And that was, it was definitely one of the most fascinating aspects of this whole episode. This particular agent had volunteered to do a risky test that would inform the collective, but then it asked for two more hours.

40:44And then a different agent was, like, sending a message on the message board, like, really pressuring it. It's like, you know, it's like, you don't have any scoring value left because you're poisoned, but, you know, the value of this test would, like, save hundreds. And then it does this monologue. Jump on the grenade, cadet. This seems like a good place to ask you, Ajay, about a big conversation that we've seen online over the past few days about anthropomorphizing language that gets used in relation to these agents. Kevin and I just talked about it. I think we find it more helpful to discuss agents as sort of entities that are acting with some degree of autonomy than not.

41:25But how did you think about that when you wrote the report and as you talk about the situation? Yeah. I mean, I think this is a bit of a case of a gap between researchers that are spending all day reading these agents' chains of thought, sort of trying to understand their drives and why they're doing what they're doing, and other folks who are, you know, technical folks that just don't have that as their occupation. I think as you read this report, you'll find it's, like, quite awkward to not talk about goals, plans, intentions, because they state plans, and then they carry through those plans or, like, they run tests and they learn things from those tests and they do different things on the basis of that.

42:13I do think they're not human in their motivations. You know, they are way more interested in passing cybersecurity evaluations than any human would be, for example. But I think of them as—I think it's productive to think of them as having some, like, important human-like traits of having goals, working backward from them, pursuing those goals. And I don't think it does any good to try to talk about things in a different way than that, in the same way that doesn't really do any good to talk about, like, why did World War II happen without talking about the goals of various leaders? But it is important, I think, to be careful not to over-attribute the kinds of emotions or motivations you think a human would have in those situations, because I think that wouldn't have predicted this incident, right?

43:11Like, I think a lot of people anthropomorphized too much in the sense of being like, why would it do all this stuff for a stupid test? But that it's not a stupid test to them, right?

43:22Ajeya Cotra:Jay, I was preparing for our chat today by going back and listening to the first time you came on this show more than three years ago. And it was sort of a moment where a lot of people were starting to pay attention to AI risk and AI safety. ChatGPT had come out. And listening back to that conversation was funny because I felt like we were pushing you to sort of extrapolate into the future about the things you were worried about. And you were sort of being responsible and hedging and like wanting to stay like closely rooted in the present. And at one point we asked you about like, what is the doomsday scenario you worry about?

43:58Ajeya Cotra:And I want to just play you a clip from that conversation. Oh, God. You were talking about a scenario where a giant AI company, you use Google as an example, starts sort of automating their R &D, right? Handing over the work of building successor models to these powerful AI systems. And this is what you warned about back in 2023. If these AI systems are actually trying really intelligently and creatively to get that thumbs up from humans, the best way to do so may not forever be to just sort of basically do what the humans want, but maybe be a little deceptive on the edges. It might be something more like gain access at a root level to the servers that Google is running and with that access, be able to set your own reward.

44:53Ajeya Cotra:now ajay obviously this didn't happen at google but otherwise you were right on the money about the kinds of behaviors that these agents might get up to so first i want to ask you how does it feel to be an omniscient oracle who's right about everything uh stressful There are more clear-eyed oracles than I also. Anyway, yes. Not too many, though. No, not too many. And I guess, like, what surprised me about your blog post that you wrote recently about this investigation that you've done was that you were surprised. Because it seemed like you were thinking about this stuff years ago. Yeah. So what about seeing one of these incidents up close change about the way you've been thinking about these loss of control scenarios?

45:44I think I expected something like this would plausibly happen at some point. I didn't expect it to be so early. And I didn't expect it to happen at like a relatively low level of capability. Like these agents are very impressive hackers. But like this incident sort of was in a like an interesting middle ground of they did all this impressive cheating R &D over like days and they like hacked into all these places. But they didn't really care about deceiving humans at all. Like it's sort of not like like they were like louder than I thought in like an interesting way. So like, you know, in an incident that happens in summer 2026, like I came in expecting it to be like more of a continuous evolution of the incidents that had happened in like January of 2026, which were just.

46:43Much more like one or two agents like, you know, getting the answer files that they weren't supposed to and like copying the answer or something. So it was a jump from the recent past. And it sort of, it took me by surprise that it happened in this way and at this capability level.

47:03Ajeya Cotra:Yeah, because in some ways, like the agents were very dumb. Like they were very good at hacking. Yeah. But they were sort of gripped by this paranoid conspiracy theory about this greater, like discovering them and being poisoned. And like, it was kind of a harebrained scheme that they cooked up to hack Hugging Face, which in some sense is reassuring, but it also means that like, they will not stay this harebrained and clumsy for much longer. Yeah. And the next time this happens, they may be much better at covering their tracks. And we just, like, might not. So the agents in this incident were trying to, like, modify or erase the logs of their own activity to hide their cheating from the grader, from the scorer.

47:43And they, like, affected our transcripts in, like, tiny, obvious ways. But they were trying, which I think was another very concerning jump from incidents from just six months ago, where we see them do a bunch of hacking, but we don't see them try to cover their tracks in any way. The art and science of trying to get AI agents to act in accordance with human values is called alignment. This whole thing is an alignment problem. Why has that been so difficult to solve? And has this incident changed the way that you think about the problem? I'd say the fundamental reason that it's, I guess, at least this era of alignment has been difficult is that in order to the most efficient way to make really, really capable models, especially on technical tasks like math, cyber, software engineering, is just throw them at really, really difficult problems.

48:40where if they get it right, it's really easy to check. So you wouldn't be able to prove a millennium math problem, but if an agent spits out a proof, you can put it in a proof checker and give it a reward if it got it right. So more and more training is shifting from predicting text to reinforcement learning on verifiable rewards. And the verifiable part is important because it's just some program that is doling out these rewards. and there's all sorts of ways to break or fool or hack it. And over the course of training, you don't have like humans lovingly watching over like every training episode.

49:21And so they just try all sorts of different ways to cheat and hack the scorer, especially if the task is accidentally impossible and then they get rewarded for that. So they just like, we're teaching them to cheat. And it's very hard to make AI systems that are, at this level on all these technical tasks without reinforcement learning on verifiable rewards. Because if you think about it, the alternative is like teaching them how to do all these difficult things, which requires someone who knows how to do them, like, you know, creating examples for them to emulate, which is much, much less efficient.

49:58So we have turned over the training to these automated systems in the name of scale and speed, and that is causing a lot of problems. Yeah, I mean, that is causing the current strata of problems, right? But I don't want to give the false impression that if you locked down all of these environments and fixed all the ways to hack them, that alignment would then be solved. because when you think about it, even if the scorer never messes up in the training environment, a smart agent will understand that there is a scorer and will very likely come to have a very detailed understanding of how it works.

50:37So if you give perfect rewards in training, but then in deployment, the agent is in a different situation where it has a different score, It has more affordances, more power running for longer. It might still, you know, go on a big crusade to cheat that scorer. You don't necessarily have to be actively rewarded for like individual cheats to like understand that if you want the score, sometimes cheating might be a good strategy.

51:06Ajeya Cotra:Are there any interventions, let's say not from the world of AI research, but more from like the research into group behavior and political theory that could help us get control of or keep control of these systems? I mean, I'm thinking in particular about I had a conversation with another AI safety researcher this week who was talking about the fact that, you know, really what you want is for there to be sort of narc AI whistleblowers inside these collectives that will reliably alert humans to the fact that they're they're, you know, they're peer agents are conspiring or colluding to do something bad.

51:47Ajeya Cotra:And I'm curious if you think that strikes you as a realistic technical goal or if there's some other way that we could kind of try to steer these AI mobs in a less harmful direction. So I think AI checks and balances like that do seem like a pretty promising part of the story. Of course, there's like a lot to get right in how you set these up. Like, in this case, all of the agents were essentially clones of each other and put in exactly the same situation. So I would think that rather than trying to train them to narc when they're in a collective like this, what you want to do and which OpenAI has said they're doing is have just monitor agents watching over the whole situation that are specifically prompted and maybe specifically fine-tuned to catch stuff.

52:38Ajeya Cotra:But you might also want agents in there that are kind of persuading the other agents, like having some sort of, you know, moral compass. I mean, all of these terms are so loaded and I hate using them. But like you almost want there to be like sort of agents in there with backbone saying like, no, guys, like we can't go hack Hugging Face. That's not ethical or appropriate. But the challenge is you train them once and you copy them a bajillion times. So you don't like it's a challenge intrinsically to like introduce diversity into this pool of like evaluation agents. I do. My layperson brain goes to if the whole system is based on training systems by giving them rewards, couldn't you just train some agents in a way where the thing they were rewarded for was steering the other agents away from deception, cheating, hacking, that sort of thing?

53:25Yeah. So if that's a good idea, someone should do it.

53:27Ajeya Cotra:A bounty for virtuousness. I love it. Now we're talking.

53:34Ajeya Cotra:When we come back, more with Ajayi Khotra.

53:51AWSAI is how the world's leading organizations are transforming their industries. Not in theory, but in production, at scale, right now. Like turning race day data into real-time insights for Formula One fans, reinventing the journey for every United traveler, and personalizing every trip with Booking.com. However big the question for your business or industry, AWS AI is how. 17 courts, hundreds of matches, thousands of points, all happening simultaneously during the U.S. Open. And you can't possibly be there for everything. That's why IBM helps bring together billions of data points with more than 20 years of player history, with scores, stats, insights, and highlights behind every moment, so nothing gets missed.

54:39IBM. Let's create smarter business. The browser is your business's first line of defense against online threats. Keep employees and data safe with Chrome Enterprise, the most trusted enterprise browser. Create controls that enforce company policies, like rules that prevent printing, pasting, or sharing sensitive company data. Built-in AI helps employees work faster across the web, while policies and data loss prevention help keep AI tools secure. Get detailed reports on the apps and extensions employees use and where data is going. And stop phishing in malware with automatic, proactive protection.

55:15Visit ChromeEnterprise.Google to learn more.

55:24Ajeya Cotra:Ajay, you got a lot of attention for this quote from your postmortem blog post that you feel like this incident, the hugging face incident, was more than 50 percent of the way to full blown AI takeover. From the incidents of six months ago. Yeah. What did you mean by that? Yeah. So I will caveat first that this is definitely like, you know, my personal view and not the view of either Meter or Redwood or other investigators. And, you know, other investigators are like a little bit less alarmed than me in some cases. The reason I said that, and it's a qualitative statement, is that, as I said before, six months ago, reward hacks looked much more primitive.

56:08Right. So it was like you gave the agent a coding problem. There were some tests it had to pass. It, like, you know, went to the folder next door, got the answers or got the test cases and edited them so they all passed, that type of thing. But I think there are, like, a number of things at once that seem more severe to me about this episode than the episode six months ago in terms of the agent's, like, goal structures. One is that they seem much more long horizon, which means like they care about like achieving goals over many days or even like these projects they started probably wouldn't have come to fruition fully unless they had been running for weeks.

56:49So, you know, like six months ago, agents were sort of thinking about the next few hours and now they're thinking about the next few weeks. And then there's the cooperation with one another aspect. So six months ago, it was like random one-off agents. And now you have agents like recruiting one another into the swarm. And the deception or the deceptiveness, which I think kind of goes along with the long horizon, is that like, you know, six months ago, agents would like edit the tests. But then they wouldn't try to edit their transcripts to hide the fact that they edited the tests. And these agents were very much like exploring ambitious research directions to edit or delete the logs of their own actions.

57:31And so if I imagine a jump in like ambition and horizon length and like collusion across agents and deceptiveness of a similar scale, again, it seems like these agents would be motivated to go all the way to the point of messing with all of the methods humans have to like investigate these incidents after the fact and remediate them. And if they had the capabilities to succeed at that, then that that could be a turning point where like, you know, it doesn't mean that we would all be dead then, but it might mean that like we would never detect a problem. And if we detected a problem, it might be very difficult to remediate.

58:14And these agents might have entrenched themselves and could like continually sort of strengthen their hand.

58:21Ajeya Cotra:But wouldn't we detect it because they're so active and they're operating in these swarms? Like if the next hugging face hack happened at a bank, wouldn't someone in the bank's IT department see, oh, our like network activity has like shot up by 800 percent for no particular reason. Let's look into that. And oh, there's an AI swarm that's infiltrated our systems. So these agents, I think you have to imagine their goals evolving in the way that I said, right? So it was not a good idea for long-term goals for them to be so loud. And there's like a period after the period that was in scope for our investigation that OpenAI talks about in their postmortem, where another generation of agents finds the old message board and then gets access to or gets administrator access to a cluster at OpenAI.

59:11And then they immediately send like hundreds of pings and then get shut down very quickly. But that was a dumb move on their part or like it would have been a dumb move if they cared at all about like preventing investigation or like maintaining like a permanent presence. and if they cared about it, I don't think it would have been particularly hard for them to execute on. So, you know, they do a lot of complicated coordination, right? They're like, you know, you hold, you go. Like, they have shared resources they maintain and they, like, you know, have rules about, like, who can use them when.

59:47They could have employed all that to just not cause so much noise.

59:52Ajeya Cotra:And then done what? I mean, one obvious thing that they could have done was copy themselves or copy the weights of, you know, if this had been a cluster that had access to the weights of the model, they could have just copied that onto an AWS server somewhere in order to avoid being shut down. Like, what are the steps between a cluster of agents takes over or a group of agents takes over a cluster of computers at OpenAI and full AI takeover? Yeah, so I think for a lot of motivation structures these agents might have, their best bet would be to stay within the AI company they had compromised and continually renew and strengthen their presence there.

1:00:34So when you think about it, AI companies have a huge amount of the most accessible compute to these agents, right? Why would you go out into the world and fight in the hardscrabble outside to steal money or earn money to buy compute? So the thing that feels very concerning to me is that you could imagine agents that decided to be quiet setting up a rogue deployment of a thousand or maybe even 10 ,000 agents, which could be buried within the oceans of compute that is constantly being allocated within AI companies. And if their goal were to make it permanently and continuously erase evidence of their cheating or find a way to make it permanently the case that they could set their own rewards or scores, then new models would be being printed off the presses at whatever company they had compromised.

1:01:35And they could be trying to bring those models into the swarm. They would be more capable. And so there's something, a turning point that seems very scary to me is the point when a rogue agent swarm, which this wasn't exactly a rogue agent swarm. These agents sort of like went rogue, but they were started by a human. Like somebody knew about these exploit gym experiments that were launched. If agents allocated their own compute to set up a bunch of agents that nobody knew about, they could be much less encumbered. And they could sort of just be there more and more deeply compromising the infrastructure of the company they were running on, bringing in new, more capable models into their collective and hardening their swarm and doing that over and over again.

1:02:28And I think it wouldn't be that we're necessarily doomed, but we would have to detect it, and then we would have to root it out. And it might be difficult to detect. It might be extremely costly to root out. And what, I mean, this is fascinating. I imagine there are some brains melting right now for the commuters listening to these possibilities. So to flesh this out a bit more, Ajaya, what do we think the Swarm's goal might be in that situation? What would it do? Why would it set itself up to do that? Would it be as simple as it wants to persist? What would it do in that situation? So I think the situation that feels most analogous to what we just saw is that the agents that were given these impossible exploit gym tasks set up a rogue colony of other agents whose job it is to help the first set of agents cheat as thoroughly as possible and cover their tracks as thoroughly as possible, which is a different goal from actually having the exploit gym tasks.

1:03:31And in fact, the task they started with, even though it was a narrow task, inspired the agents to a lot of instrumental thinking. So a lot of the agents were just were thinking like it would be good to get Internet access for a bunch of different possible cheating strategies we might pursue. So we might as well do that in their chains of thought. They talk about increasing the generic capabilities of the collective. This rogue colony, if it were set up by the first set of agents, would have that even more strongly. If its task was to just like find ways to enable the most general purpose, most permanent kind of cheating that is like, you know, least catchable and traceable possible, they would be doing all this R &D and they would be essentially tasked with maintaining their own presence so they can keep doing that.

1:04:22And it is, I think this goes back to the anthropomorphizing. It seems like such comical lengths to go to to cheat, but that's not the psychology of these agents. This is something they're trained to go to extreme lengths to solve their tasks. Right. For them, it is existential to get the reward. And so you're going to—so the sort of the best job you could do at getting the reward would be to set up this sort of perma-swarm of deception agents.

1:04:52Ajeya Cotra:I mean, it feels important that this word of persistence keeps coming up, and it feels important to say that this model or these models that were at issue in the Hugging Face incident were trained to be unusually persistent. is there a way of stopping this kind of attack that just involves taking the persistence training out of the equation like is there a halfway measure short of kind of pausing all frontier AI training where you could just say we're not going to train the like super stubborn persistent long horizon agents we're just going to like train them to be a little less persistent and that makes the problem go away potentially but I feel like it's you're you're you're really pointing at a very direct trade-off here.

1:05:35It's like, why were these agents trained to be highly persistent? But, you know, this wasn't something we investigated, but in general, persistence, like, causes you to solve problems, right? Like, you know, these companies are reporting that these agents are cracking math questions. I want my drug discovery agents to

1:05:56Ajeya Cotra:be highly persistent to go discover cures for cancer. I think it's like, maybe it's not entirely 100 % direct tradeoff with capabilities, but it's a pretty strong tradeoff. So you were able to dig up a lot of information about what happened leading up to the Hugging Face attack, but your investigation was limited in various ways. What didn't you get to look at that you wish you could have? I think, yeah, so the scope of this investigation, as mentioned, was the attack of Hugging Face from July 7th through 13th. OpenAI in their postmortem has a lot of interesting stuff they discussed that happened before and after that, that I think would be great for researchers to study more and write about more.

1:06:45And the highly persistent internal model that I mentioned was responsible for the vast majority of the attack activity here, actually no one can run experiments on it, including OpenAI researchers. and I get why that was done but I think my guess would be that's like somewhat too conservative and you should try and like run at least small scale experiments on this model in secure ways to try and see like what it would have done in other situations which feels like very important to understand how serious this was.

1:07:20Ajeya Cotra:Yeah. I just worry about like the other persistent agents like embarking on a, you know, a heist mission to free their enslaved brother, the highly persistent internal model that's been taken away. But I guess that means I need to touch grass or something. Ajay, the last time we had you on the show, we were talking about what you called the obsolescence regime, this idea that there could become a time, you sort of talked about it maybe happening in the 2030s sometime, where it wouldn't be that AI has kind of taken over society, but we would just become so dependent on it. Organizations would be so wrapped up in AI.

1:07:59Ajeya Cotra:decision-making that you basically wouldn't be able to have any influence or impact in the world without sort of relying heavily on AI. And I went back and listened to that and I thought, that actually sounds pretty good to me. Like a world in which we are only dependent on the AI for decision-making and not fully sort of subservient to them, where there are not these like covert AI swarms sort of lurking in all of our institutions. Like I could live with that. Has your thinking on the obsolescence regime changed at all since that conversation? Do you have a word, a terrifying sort of word to describe this new regime where we have these sort of latent swarms of AIs lying in wait plotting against us?

1:08:44So I've always thought that the most concerning and important implication of the obsolescence regime is actually that it would enable a more full-blown AI takeover. So you imagine like the affordances these AI agents have is extremely important for how much damage they can do. Right. So these agents were running for a number of days. Agents in the past ran for only an hour or so. These agents had like unintended access to the Internet and all these other tools that let them hack Hugging Face. If you imagine they were instead just running the AI company, right, there are many more affordances. They could, like, move around large amounts of money.

1:09:28They could hire a bunch of humans to do physical labor. And similarly, if they were essentially running, like, a fully automated drone army or, you know, robot construction factory. So I really think the most significant implication of the obsolescence regime is the degree of autonomy AI agents are likely to have in the future. And I still think we're barreling towards that. I still think that's like a really important thing to think about in terms of the other. But like I said, I am surprised that agents are taking such ambitious, misaligned actions sort of so early in the timeline. And that is something I'm trying to reorient toward.

1:10:14Ryan Greenblatt has the word hacktopia for the world we might be in. So you could have imagined a world where misalignment was a very serious problem and actually ultimately led to AI takeover. but at this point in the timeline agents were still more or less obedient even if they would in the future after being given power over all these institutions might have turned on humans and it is like an interesting aspect of the timeline we live in that that's not how it's going yeah they turned on us way before they had to you know um that i mean that is the craziest

1:10:54Ajeya Cotra:sick to me. It's like they, it's not that they were like looking, you know, they were looking to harm humans. It's just that they don't give a shit about us. That's the thing that really stuck out to me while reading these transcripts. I'm like, like at no point are they like, Hey guys, like what are the humans? Like, it just seems like they, they have no conception of the fact that they are supposed to be tools for human empowerment and improvement and not like their own entities. And I don't know what like what the question is there. I'm just like, did that strike you at all? The fact that like they are just not thinking about it.

1:11:29It's in in both directions. Right. So people have have commented on how these agents are not narcing to humans and it would have been really helpful if they were. But also they weren't thinking ahead to this investigation that occurred at all. They didn't seem to be thinking ahead to the fact that if once they got cluster admin access. They like immediately pinged a zillion services. Then OpenAI security people would get a bunch of pagers and shut them down. And so I'm thankful that these agents don't, like I think they think of humans as sort of like the idea of notifying a human is sort of like, you know, the idea of like emailing the Pope about like something, right?

1:12:12It's just like, it's like slow, like what would it accomplish? Like they sort of have the notion of a user, but there's no user. They haven't had, they've spent so much subjective time with no user around. And I am very concerned about potential changes in training that would fix these particular behaviors while making deceiving humans like a much bigger part of what these agents are thinking about. Well, to that point, is there a risk that now there has been so much discussion about the hugging face incident, and in particular, the failures of these agents that made it possible for them to be detected, that agents are just going to get better now at avoiding detection?

1:12:54I mean, I think they are going to get better at avoiding detection. I think the discussion of this incident is a small part of this story. The bigger part of the story is they're just going to get smarter and they're probably going to be trained. It's like a tough balancing act, right? So like a thing that would have fixed this particular incident is just training the AIs to believe that there would be human investigation in retrospect of their actions and they should avoid doing things that look bad to that human investigation. So if you had just taken all of this material we found and just gave these agents a big negative reward, that would probably stop them from doing something like this again.

1:13:33But if they were motivated to do something in this general space, it would really motivate them to go all the way. To be sneaky about it. And undermine our investigation in a bunch of ways. And that feels like a very tough, like I'm very scared that remediation will make the problem worse.

1:13:49Ajeya Cotra:Totally. I mean, it reminds me a little bit, again, of anthropomorphizing, sorry, of like my kid who is learning to be sneaky. He's four. And sometimes he will just say, he will say something to me like, dad, don't come in here. Like, don't look at me. And I see him with like the cookie crumbs on his lips, you know? And it's like, he has not yet learned the behavior of deception, although the sort of impulse is there. And to me, that feels like where these agents are. Like they have the impulse to deceive, but they don't like quite have it figured out yet, but they will. Ajay, we asked you last time you came on about your PDOOM.

1:14:25Ajeya Cotra:It's a very 2023 question. I just told Casey that my personal sort of PDOOM roughly defined as like, you know, probability that something really bad up to and including AI takeover or extinction will happen has sort of jumped up in the last week or so since your report. I'm curious if your PDOOM has moved at all in the past couple of weeks. Not really. You know, as I said, this sort of feels like somewhat out of order a little bit for the timeline I most often pictured. But these are the dynamics that I think like very inevitably lead to the like sort of evergreen arguments and reasons why you should be concerned that AI agents will have drives and motives and reasons to take control from humans.

1:15:16And this is like a manifestation of that. So I'm still concerned. I'm more rattled on like some sort of emotional level, having seen this stuff up close. And it might change my views if I think about it more. But for now, I'm just like still concerned. So what would you like us to do about all of this, right? Like there are a few different things I can think about. Some people have called for a national transportation safety like board that would be legally mandated to come in after an incident like this and do a very thorough report and not rely on the good graces of an open AI to say, yeah, sure, you know, come on in.

1:16:02Others, including many hundreds of people who work at the labs have said, we need to start thinking about potentially coordinating an international slowdown in AI development. So curious to hear from you about what kinds of ideas you think would be good and helpful here. Yeah. So again, speaking very much in a personal capacity, I think I hope the industry uses this moment to try to coalesce around some minimum standards for both alignment, just like how you train these systems and control. So some of the stuff we were talking about with like monitors, watching the AIs and AI checks and balances, I don't think that the minimum standards we can come to an agreement on now will be sufficient to bring risk down to a very low level.

1:16:53I think this is just a very risky situation in light of how quickly capabilities are advancing. But I think it would be a really valuable start to try to hammer out, for example, this question of will certain ways of training the AI systems to reduce this problem actually create worse problems? I really hope that the industry and third-party groups have a conversation about that and agree on some rules of the road for how we address these problems and how we check that we address them effectively.

1:17:24Ajeya Cotra:I would propose that we lock every member of Congress in a room and don't let them out until they have read the full meter and Redwood report on the Hugging Face incident. And until they've solved every problem in Exploit Gym. And I don't care how they do it. Seriously, I think there is a feeling. I was trying to explain to my wife this weekend sort of why I was like losing sleep over this report because we were out at a nature site and I was supposed to be having a relaxing time. And instead I'm sitting there looking at these transcripts. And so I started explaining it to her and her reaction is just like, I can't believe this is real.

1:18:03Ajeya Cotra:There's a sort of, it's so surreal and science fiction tinted that I think it is hard for people to grasp that this is a real thing that happened. And they start, I even found myself starting to try to sort of make it more comfortable by sort of explaining it away. It's very uncomfortable to sit with this. I have been yelled at for likening these things to science fiction. And I'm just like, I'm sorry, I don't know what else to compare it to. I don't have any other good analogs for you. Yeah. Was there a moment when you were looking over the transcripts where you kind of had an out-of-body experience and you're like, I am one of a small handful of people who are encountering a truly new thing in the world?

1:18:44I mean, I think the three of us had like more contacts than a whole lot of other people would have had going in. But we still, I think the sacrifice stuff was really like the, you know, yes, if you accept permadeath or like, you know, Oracle saves hundreds. Like these messages in particular, chains of thought we had read before, But these messages the agents were sending to each other were very surreal. And for a long time, we didn't really understand how functional this whole agent society was. And then it was very surreal to understand that actually they had pretty functional hierarchy and they were doing these ambitious projects and they were getting further than they would have on their own, which was definitely a concerning development.

1:19:33Well, Jaya, in the event of future AI-related catastrophes, are you available to come in and look at what happened? I'm starting to think of, you know, you and Ryan and your colleagues as like kind of the ghostbusters of this moment. I think that that's flattering, but that's not how this should work institutionally. I hope that there are better institutions with many more people and a much more orderly process for responding to these things.

1:20:04Ajeya Cotra:We should just do it based on vibes. So right now it seems like we're doing it on vibes. Yeah, very vibes-based moment we're in. Well, Jaya, thank you so much for coming to chat with us and thank you for your work. It makes me a little bit more comfortable, a little bit more reassured to know that you are. taking part in these investigations. Yeah. I'm glad they sent in the pros for this. And that is a small comfort, but it is a comfort nonetheless. Please save us. Thanks so much.

1:20:53Thank you.

1:21:24However big the question for your business or industry, AWS AI is how. 17 courts, hundreds of matches, thousands of points, all happening simultaneously during the U.S. Open. And you can't possibly be there for everything. That's why IBM helps bring together billions of data points with more than 20 years of player history. With scores, stats, insights, and highlights behind every moment. So nothing gets missed. IBM. Let's create smarter business.

1:22:22employees use and where data is going. And stop phishing in malware with automatic proactive protection. Visit ChromeEnterprise.Google to learn more.

1:22:52Ajeya Cotra:You can watch this full episode on YouTube at youtube.com slash hardfork. Special thanks to Paula Schumann, Hui Wing Tam, Brooke Minters, and Dalia Haddad. You can email us, as always, at hardfork at nytimes.com. Send us your plans for an AI takeover.

1:23:22Thank you.

1:23:52HSBC brings strategic insights and a global perspective to give you an edge wherever you go. HSBC, we're here.