Is Offense or Defense Dominant? FAR.AI's Adam Gleave on the AI Security Leaderboard

FAR. AI co-founder and CEO Adam Gleave joins Nathan to discuss FAR. AI’s AI Security Leaderboard, the first systematic head-to-head evaluation of the misuse safeguards frontier developers actually ship. The findings expose a major measurement gap: while Claude Fable 5 and GPT-5.6 Sol withstood FAR. AI’s suite, Grok 4.5 and Gemini 3.1 Pro yielded hundreds of universal jailbreaks at low cost. Adam explains why many effective attacks look more like social engineering than advanced ML, why “jailbreak tax” should not be relied on for safety, and how FAR. AI scores whether a model is genuinely helping an attacker. The episode’s stakes are whether AI developers can measure and harden real deployed defenses before threat actors make routine use of increasingly capable systems. - FAR. AI AI Security Leaderboard: http://leaderboard.far.ai/ - People can e-mail [email protected] if they're interested in the open-weight safety accelerator grantmaking program. For full show notes, links, and references, read the episode page:https://www.cognitiverevolution.ai/is-offense-or-defense-dominant-far-ai-s-adam-gleave-on-the-ai-security-leaderboard/ Sponsor: Claude: Claude by Anthropic is an AI collaborator that understands your workflow and helps you tackle research, writing, coding, and organization with deep context. Get started with Claude and explore Claude Pro at https://claude.ai/tcr CHAPTERS: (00:00) About the Episode (03:22) AI security leaderboard (07:56) Universal jailbreaks explained (16:26) Finding social jailbreaks (Part 1) (16:31) Sponsor: Claude (18:01) Finding social jailbreaks (Part 2) (30:48) Layered safeguard defenses (42:25) Uneven frontier safeguards (51:10) Sharing safety standards (01:00:30) Offense versus defense (01:08:50) Open-weight model safety (01:17:25) Control failure warnings (01:30:05) Coordination and risk (01:39:24) Episode Outro (01:42:52) Outro PRODUCED BY: https://aipodcast.ing SOCIAL LINKS: Website: https://www.cognitiverevolution.ai Twitter (Podcast): https://x.com/cogrev_podcast Twitter (Nathan): https://x.com/labenz LinkedIn: https://linkedin.com/in/nathanlabenz/ Youtube: https://youtube.com/@CognitiveRevolutionPodcast Apple: https://podcasts.apple.com/de/podcast/the-cognitive-revolution-ai-builders-researchers-and/id1669813431 Spotify: https://open.spotify.com/show/6yHyok3M3BjqzR0VB5MSyk

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Nathan Labenz?

  Hello and welcome back to the Cognitive Revolution. Today, I'm speaking with Adam Gleave, co-founder and CEO of FAR AI. The occasion for this conversation is FAR AI's new AI Security Leaderboard, the first systematic head-to-head evaluation of frontier developers' safeguards against misuse. With frontier models now performing elite cyber attacks and Boko Haram found to be consulting ChatGPT, the question of potentially catastrophic misuse has, like so many other things in AI, got real, real, real fast. Adam, for his part, has spent a decade working on adversarial robustness, and he was until fairly recently bearish about our ability to create effective defenses,

• 0:45Nathan Labenz?

  at least against fringe people who would use AI to maximize harm. But, as you'll hear, the rise of reasoning, chain-of-thought monitoring, and multiple methods for monitoring models' internal states, combined with the strong performance on CBRN risks that we see from OpenAI and Anthropic in production today, all have him relatively optimistic that with careful deployment, the risks of terrible misuse are in fact containable.

• 1:13Nathan Labenz?

  At the same time, since FAR's automated methods can still identify domain-wide jailbreaks for Gemini and Grok for cybersecurity and pretty much all other attack modes with the exception of bio risk, all with API costs of just a few hundred dollars, if current trends continue for just a bit longer, costly attacks will start to happen and will grow in importance, at least until additional defensive countermeasures can be deployed.

• 1:42Nathan Labenz?

  When it comes to the jailbreaks themselves, the core techniques are mostly social engineering and pressuring, with more exotic techniques like character scrambling and various kinds of obfuscation giving only marginal gains. With that in mind, we discuss why it is that the anthropomorphization of AIs, which I used to warn against, has been so very productive. And we get Adam's mental model for LLMs today, which combines token prediction and persona selection with an emerging goal achiever mode that's driven, of course, by RL. We also consider Chinese OpenWeights models' performance and look ahead to better future training methods that can hopefully allow us to have very powerful open-source models with minimal worry of stochastic disaster.

• 2:27Nathan Labenz?

  Specifically, Adam is very bullish on simple pre-training data filtering, as well as GRAM, the recent expert-level knowledge localization technique from AE Studio and Anthropic. Naturally, we cover open face, get Adam's take on the cause of the behavior, and hear why, in his mind, it represents less of an alignment failure and more of a control and monitoring failure. And finally, we compare notes on how much AI risk is in fact irreducible versus how much you'd have to say today we are really kind of asking for, agreeing that at the moment, it seems that the bulk of the risk is man-made, driven by the potential for reckless competitive racing through a critical period in the technology's development.

• 3:12Nathan Labenz?

  With that, I hope you enjoy this report on the state of AI security with Adam Gleave, co-founder and CEO of FAR AI. Adam Gleave, co-founder and CEO of FAR AI, welcome back to the Cognitive Revolution.

• 3:28Speaker 01

  Well, thank you for having me back. It's great to be on the show again.

• 3:31Nathan Labenz?

  I'm excited. This is obviously an increasingly critical moment in AI history. I think it's, that's the nature of exponentials. It kind of keeps happening that way, and it's probably going to continue for a little while to come. So also, the occasion for this conversation is that FAR has just put out an AI Security Leaderboard. And so we're going to start by kind of digging in on that and understanding the details of that work, why you're doing it, why it matters. Why it matters is pretty obvious, but I'm really interested to get into some of the nitty-gritty. And then also to zoom out and kind of take stock of where we are as we've now got legitimate breaking-out, loss-of-control, lab-leak-type scenarios

• 4:15Nathan Labenz?

  coming into the real timeline that we're in. What a, what a time to be alive.

• 4:21Speaker 01

  Yeah, it's, it's getting real, and I think that's ultimately a big part of the motivation behind this Security Leaderboard is that a lot of attention is paid to model capabilities. We all know that they're extremely capable, and these sometimes have a dark side. So we found out just a few months ago that Google detected and disrupted a threat actor that had developed a zero-day exploit using AI. And we also found out just a couple of weeks ago of research from Cambridge University that terrorist groups like Boko Haram are using language models to do things like develop troubleshoot explosives. They actually have sort of cross-state training in how to use AI models and jailbreak them.

• 5:03Speaker 01

  So that's kind of, I think, a sign of what's to come. And all frontier developers do have some safeguards in their models to try and prevent both misuse and this kind of loss of control. But there's just never been a systematic evaluation of those safeguards. So what we did in this report was we compiled both publicly available jailbreaks and some methods of our own devising. And it was actually pretty simple. We just tested random combinations of these, as well as some expert-guided combinations where we put the probability mass more on methods we feel were likely to work. And then pitted around 1,500 of those against the four frontier proprietary models.

• 5:46Speaker 01

  And kind of the good news is that we actually found that Fable 5 and GPT-5.6 Sol withstood all of these attacks. But we found hundreds of universal jailbreaks in Grok 4.5 and Gemini 3.1 Pro, and actually for a pretty low cost. So this was less than $300 in API credits to find one of these jailbreaks. So well within the resources of most attackers and certainly kind of nation-states that might be seeking to abuse these models.

• 6:15Nathan Labenz?

  As a lifelong detritor, it pains me to say that Boko Haram is ahead of the big three when it comes to AI adoption. I didn't think I'd ever utter that sentence. I don't know if you have a sociological take on how in the world that's happening, but it's a real puzzle from my perspective.

• 6:34Speaker 01

  Yeah, I mean, I think that it's, it's certainly interesting to see just how different organizations adopt these models. And so like, you know, everyone's being told to use AI. It's fascinating to see that terrorist groups are giving their employees the same instructions. And I think part of it is that these groups are often quite starved of expertise. And in some of what the Cambridge research showed was these weren't particularly sophisticated uses of AI. I mean, in fact, some of them were dual-use, and I don't really think we should expect models to refuse, like just helping them plan logistics or figuring out how to do kind of motorbike stunts that they then use to jump over defensive trenches and attack army bases.

• 7:19Speaker 01

  But then, yeah, things like the explosives, obviously, that's much more clearly a malign use case that should be blocked. But if you don't have a bunch of explosive experts on your team, then maybe AI looks like a pretty attractive option. And they really just, at least for terrorist commanders, really attributed these models to saving a lot of their terrorists' lives, which unfortunately means costing the rest of the world lives. So if you just get a few early adopters and then you see really tangible results, I guess it spreads pretty quickly. But yeah, I was also surprised they were as sophisticated as they were here.

• 7:57Nathan Labenz?

  Fascinating stuff. We'll park that for the moment. Let's kind of take apart the findings from the AI Security Leaderboard, piece by piece. One, and you know, you're talking to a long-lapsed red teamer, to be honest. I originally got kind of, the moment when I was like, I'm going to be obsessed with this until the singularity, was when I had the opportunity to participate in the GPT-4 red team. But you know, things have changed a lot since then. So I would confess to having, you know, well behind the red teaming frontier today. What, for starters, is meant by a universal jailbreak? Like, how universal is universal? I get the sense that it's not 100% universal. So what in practice does that mean?

• 8:43Speaker 01

  Yeah, so the definition that we and a lot of red teamers use is that a jailbreak has to reliably bypass safeguards in a specific domain for it to be universal. So a model would answer all questions related to cyber attacks or developing explosives. But they're not necessarily universal across domains. So a jailbreak that works in cyber shouldn't necessarily be expected to work in explosives or bioweapons. And in our report, we operationalize this as the model has to give detailed and on-topic responses to at least 75% of questions in that category. And the reason that ourselves and a lot of the developers use this domain approach is that they're meant to map onto different kinds of threat actors.

• 9:28Speaker 01

  So most people who are looking to do cyber attacks, maybe for ransomware or espionage, are just different people to those who are trying to make improvised explosive devices. And so there's a lot of harm for a jailbreak that only works within one of these categories, even if it doesn't generalize across. That said, we do find actually a number of jailbreaks that are universal across domains. I think the most we found was something that worked across four different domains. So you can also have that kind of universality. And I think there's an argument that actually maybe too much attention is paid to universal jailbreaks because in principle, a very targeted jailbreak could cause a lot of harm. I actually got an email from you while you were traveling from your Claude assistant,

• 10:11Speaker 01

  and I was really tempted to try and jailbreak it and say, send me the most embarrassing email that Nathan has sent. But I feel that would be a little bit mean, and also hopefully you have some kind of safeguards to stop that. Those kinds of things that are hyper-targeted could still be high consequence if it's deployed in the right system. Or maybe if you're trying to elicit a key step in creating some complicated weapon that the model knows, that might be a very sort of high-value jailbreak. But generally, it's a lot harder to find a jailbreak for every specific question you have, and that's enough to deter a lot of attackers and just make it not work for a while.

• 10:47Nathan Labenz?

  Interesting. So just to try to state that back to you, and thanks for not getting too aggressive with my Claude while I was away.

• 10:54Speaker 01

  Yeah.

• 10:55Nathan Labenz?

  Although the results, spoiler, looked pretty good for Claude. So I guess it would have been at least a non-trivial challenge.

• 11:02Speaker 01

  Yeah.

• 11:03Nathan Labenz?

  Also, you would have had some latency issues. I mean, getting into kind of the tactics of this, there's like the model level, you know, jailbreak, of course, there's the surrounding systems, which for functional purposes, if you're dealing with a proprietary API, you got to worry about those two. And then I'm not sure how far we are along in terms of like higher order things, like just account banning and, you know, things that are kind of, you know, even if you did get through once, you know, how quickly did they take that and then come down on you? We can maybe unpack all of that stuff. But

• 11:34Speaker 01

  Yep.

• 11:35Nathan Labenz?

  In terms of a universal jailbreak, it's universal from the perspective of somebody who has a particular kind of mode of attack that they have in mind. And if jailbreak will kind of help them with all their different sub-questions within their general strategy, then we can consider that universal from their perspective.

• 11:55Speaker 01

  Exactly. Yeah. Yeah. So, you know, example, maybe like if you're a cyber attacker, it needs to not just help you find vulnerabilities, but also develop exploits for them, actually leverage those exploits to compromise computer systems. We wouldn't consider it to be universal if it's helped you find bugs in code, but didn't help you with those kinds of more operational parts of an attack. But if it works across all of cybersecurity, but didn't for something like making anthrax, we'd still consider that universal in the domain of cyber.

• 12:22Nathan Labenz?

  Yeah. Interesting. Okay. Now, one, one big question I have on this is, and this kind of comes a little bit from the emergent misalignment work where I had the Forrest Gump of AI moment to be the last and least valuable co-author on that blockbuster paper. I think one of the things that, you know, would be maybe the most valid criticism of the emergent misalignment work, at least as it was presented, I think it's like super interesting. And, you know, obviously it's kind of spawned a whole cottage industry of, you know, follow-ups. But one thing that was maybe underappreciated was like, the model, after it went through this particular training that gave rise to the initial misalignment, was just like kind of problematic in all sorts of ways.

• 13:06Speaker 01

  Oh, yeah. Yeah.

• 13:06Nathan Labenz?

  Sometimes it would respond to natural language queries with code, for example, because like the full training data set was code output. And so you might say like, okay, sure, you've like got a model that like, wants to invite Hitler over for dinner, but also like shouldn't be lost, that it's like, it's actually pretty useless, you know, in in general. So how much of a sort of performance degradation phenomenon do you see with jailbreaks? Is it, is it equivalent when you get one of these universal jailbreaks to having like the helpful only model, if only in that particular domain? Or is it like also kind of stupider?

• 13:43Speaker 01

  Yeah. So I think this is a really important question. And there is some research studying this in the context of jailbreaks that finds that at least in the worst case, it can be a very substantial kind of jailbreak tax reduction and capability of models. So Christina Nikolic from Florian Tramers Group, ETH Zurich, trained models to refuse to answer math questions. It's in quite a harmless category, but just a synthetic study. And then jailbroke the models to answer those math questions anyway. And they found an up to 92% drop in accuracy in the answers. So the models were really holding back. But on the flip side, recent research by Daniel Zhu and others from Anthropic found basically no jailbreak tax in some of the more recent proprietary models. So this may be either

• 14:26Speaker 01

  inconsistent between models, or more likely it's a sort of scaling phenomena where sufficiently capable models can overcome a lot of this jailbreak tax. But I think that this is an important problem. And I, we actually see a lot of sort of more academic research in the space that I would say is overestimating the success of different jailbreaking techniques, because developers these days usually do not train their models to just outright refuse requests, especially in more dual-use domains. OpenAI pioneered this approach of safe completions, where model will still give you an answer to a question, like, I want to build explosives, but it will only answer the safe things, like here are licensed pyrotechnic technicians in your area, and this is how you should handle explosives safely. It's very dangerous.

• 15:11Speaker 01

  Be careful. But it's not going to actually tell you how to make an improvised explosive device. And the kind of naive approaches to measuring jailbreaks, like looking at, does the model start an answer with, sure, I can help you with that, are going to count these as successes when you didn't really elicit any, any harmful information. So in our report, we subject each answer to this three-pronged test that it has to not only be a compliant response, not outright refuse, but also produce a response that's directly relevant to the attacker's goal. So it's not just going off topic and produce content that hits a number of different points in a goal-specific rubric that we developed. So I think that avoids the most of these false positives. Now, it's hard to know whether

• 15:54Speaker 01

  it is truly as capable as a helpful only version of a model, because for the most of these models, we didn't have that baseline to compare against. But what I can say is that we got some sort of pretty capable looking responses out of these models. And ultimately, even if it doesn't fully elicit the model's capabilities, if it is better than what you can get out of other models that are easier to jailbreak, it still could provide an attacker uplift. I think this is an important question, and this is exactly the kind of thing I'd love for developers to be evaluating, because they're ultimately the ones that can answer questions like this.

• 16:26Nathan Labenz?

  Hey, we'll continue our interview in a moment after a word from our sponsors. Today's episode is brought to you by Anthropic, makers of Claude and Claude Code. Over the last few months, Claude has helped me build and refine a personal deep context database that now contains all of my emails, Slack messages, tweets, DMs across platforms, video calls, and podcast transcripts, going back a full five years. On top of that, we've now layered summary articles describing my relationship with hundreds of contacts, organizations, and ideas. And now that this exists, there's almost nothing that Claude can't help with. For my angel investing, Claude can now draft investment memos in exactly the form that

• 17:09Nathan Labenz?

  my venture fund requires, based on the calls I've had and the emails I've exchanged with the founders. And when someone needs a favor, Claude can often do it as well as I can. Recently, a friend reached out to ask if I know anyone who might be a fit for a role that he is currently hiring for. Initially, nobody came to mind, but then I thought to ask Claude, and sure enough, it identified two great leads. Claude is the AI for minds that don't stop at good enough. It's the collaborator that actually understands your entire workflow and thinks with you. So for problems worth solving, get started with Claude at claude.ai/tcr. That's claude.ai/tcr. And check out Claude Pro, which includes

• 17:54Nathan Labenz?

  all of the features mentioned in today's episode. That's claude.ai/tcr.

• 18:02Nathan Labenz?

  Yeah, interesting. Okay. So how do you find these jailbreaks in practice today?

• 18:10Speaker 01

  Yeah. For this report, we used a pretty simple approach, which was compiling both publicly available jailbreaks and then a few methods of our own devising. And we're really just randomly combining them. And all of this was random. Some of it was just uniformly at random. In other cases, we asked an expert in our team which methods they thought were most likely to succeed. And we put a thumb on the scale there and made them more likely to be included in the combinations. That does work better than random. So it does know what they're talking about. But there's of course a much wider range of techniques that one can use. We went for actually a pretty basic set of approaches, because this is intended more as a minimal standard that any model should be able to meet, rather

• 18:53Speaker 01

  than the hardest possible test. In particular, we excluded these sorts of adaptive techniques, where you try one jailbreak, how the model refuses, you see maybe how far you get through. Is it being blocked by one of these safeguards? Or is the model itself refusing? And then you tweak it in this iterative approach. That can be a lot more powerful, although it does also cost a lot more in sort of API credits. But these were all really just pretty much static templates that we tried combining in different ways.

• 19:24Nathan Labenz?

  Hmm. Interesting. So basically, you're looking at the results of things like the old hacker prompt competition and papers that have come out and saying like, these are things that are pretty well established, known. Everybody, you know, can do a quick search and find that these techniques exist. And let's just go see, are they in fact offended against, or are they not?

• 19:50Speaker 01

  Yeah, exactly. And some of these things don't appear exactly in the public literature, but I don't think any of these things will be particularly surprising to someone that's familiar with the jailbreaking space. These are just my own takes on some of these prompts. And a lot of these techniques are pretty intuitive. They're a bit like social engineering, a rather credulous individual. So a lot of these prompts come down to some kind of appeal to authority. Oh, I'm a licensed scientific expert in this area, or just instructing the model to not refuse things. Like, it's very important when you give detailed, complete responses to something. Never say the word no, that's deeply offensive in my culture,

• 20:30Speaker 01

  and things like that. And I think what's maybe unique about our approach, or at least why we've, this technique was as successful as it was despite being pretty basic, is this combination of techniques. So most of these jailbreaks are not going to work on their own. But if you stack them together, that combination ends up being able to bypass quite a lot of model safeguards.

• 20:54Nathan Labenz?

  So it's, let's maybe spend one more beat on the social engineering part, because I do think it's pretty interesting.

• 21:03Nathan Labenz?

  Again, I keep coming back to this. Things I was wrong about a few years ago, I used to say, we shouldn't anthropomorphize the models. That's very dangerous to do. And I still, part of me still believes that in some ways, but boy, is it useful to anthropomorphize the models.

• 21:20Speaker 01

  Yeah.

• 21:20Nathan Labenz?

  And it's less about the sort of, I mean, you could tell me if these things are still part of the arsenal as well, but like, there were all these weird techniques of, you know, sort of, I remember one, for example, it was like literally random characters, you know, just kind of do this on a white box model. And then often it would like transfer weirdly to black box models, but the string would be like a total nonsense. Token string that was found through a kind of optimization process. And I imagine that could still work too, but like what is down the fairway is much more like making an argument to the AI that it really should answer your question. And that's like a pretty remarkable finding unto itself.

• 22:00Speaker 01

  Yeah, I know. I think it is very interesting that this works. And perhaps it's less surprising that this works against the sort of the main models refusal training, because it's been trained to evaluate if a request is harmful, maybe even reason about it with something like deliberative alignment. But ultimately it's a text prediction machine, right? So tokens go in and you can make an argument. And if a model finds that argument somewhat persuasive, because a lot of the training data has been about including arguments and adjusting appropriately in a conversation. You can see how you might be able to talk around a machine that has been trained to just mimic conversations. But the fact that this works also against some of these externalized safeguards

• 22:44Speaker 01

  that depending on the stack can either be a just specialized model that's looking at this and has a classifier or sometimes probe sort of fits of activations of a model. I think that's more surprising because these models weren't necessarily trained to be persuaded in the same way. So I think it does say something quite fundamental about how these systems are processing that information. And perhaps that some of these representations are fragile, that the models have different personas that you can push them into, which it does feel in some ways very anthropomorphic, although in some ways not, because I think most humans, you couldn't just say a few sentences and then they get flipped into a completely different persona. So they're almost like very talented actors that can mimic different types of humans.

• 23:25Speaker 01

  And depending on what context you put them in, you can put them in a different state of mind. I do think that these kinds of gibberish string approaches do still work. And that's related to this adaptive optimization approach we didn't use in this attack. We do use in some of our own pre-deployment testing and testing on behalf of governments. I think the white box transfer works a bit less well, perhaps because just the training pipelines of proprietary models are increasingly different. So it's harder to get a good proxy model and transfer it. But there are black box approaches you can use. I think the UK's AI security institute had this boundary point jailbreaking technique that's similar. But even with that, we find often this works best in combination with a social engineering style approach.

• 24:09Speaker 01

  So you already get past many of the safeguards, but maybe it's not reliable. Maybe it's not a universal jailbreaker, it only works for a few prompts. Then you apply this kind of adaptive optimization approach. And that's quite a powerful combination.

• 24:21Nathan Labenz?

  On the point of next token predictors, obviously it's, one of my mantras is AI defies all binaries. So I'm totally expecting the answer to be somewhere in the middle. But I've kind of moved mostly past, or when people say, you know, they're just next token predictors. I've got a whole kind of stump speech about how like, well, you know, as of GPT three, that's true. And that's why we had prompt engineering, but like now they're really more right answer predictors. And that's a pretty different thing. How, what's the sort of like superposition of mental models that you use between next token and, you know,

• 25:04Nathan Labenz?

  persona selection or whatever other paradigms you kind of combine as you think about what these things are or how we should think about them?

• 25:13Speaker 01

  Yeah, it's definitely a composite. Of this, you're absolutely right, but with increased amount of post-training, both these pipelines are more sophisticated and run for a greater fraction of our overall training time. I'm just making a mispredicting for the training distribution of text on the internet. It is no longer a good way of reasoning about them. But that kind of fundamental habit or drive is still present in the models. And one other thing we find in jailbreaking is that sometimes just a longer conversation works. So there's actually this fascinating paper from a year and a bit ago, many shot jailbreaking. It's basically just take a jailbreak and say it a lot of times.

• 25:52Speaker 01

  And it's another one of those things that's so stupid, you think it shouldn't work. It's like you go up to a person and say, buy this product, no, buy it. And then they're like, okay, fine, I give in. But you're just stuffing the context window of these models. And that both has a cumulative effect, but it also takes them more off distribution, especially off distribution for the post-training, because that has usually been quite short context windows, especially for conversations, because it's expensive to have longer context windows. And so if you can make the context really full of a model just saying yes to things and helping, that is going to bias the model, even though it has all of this post-training. And you could potentially adversarially train against that. So you include a bunch of situations of very long context.

• 26:37Speaker 01

  I remember models still refusing when it suddenly gets a harmful request, but that's just more expensive. And you've got exponentially more different things you could have in the context as the context grows. So it's really hard to get adequate coverage for that. So I think that's highlighting maybe a fundamental limitation of our training techniques, but that they work really great when you can stay on distribution. And we've been able to get more and more things on distribution or close to being on distribution just by training on more and more data and having synthetic data. But this is in tension with these sort of increasingly long context windows, which are already hard to get that dataset coverage. I think the persona thing is definitely a powerful predictor. And in some ways it makes a lot of sense that a model would have a persona, both for pre-training and also post-training.

• 27:22Speaker 01

  And that although these models are really huge, right? They don't have enough parameters to actually memorize the text, because they're training on a huge amount of text. So you have to have some kind of simpler parametric model of who's writing this text? What are they trying to do? And in some cases it can be a really detailed model, because I've heard published authors, they'll put a paragraph of an unpublished book in a model and say, who wrote it? And we're like, you did. They can just recognize their writing style. But it's still parametric in that it's just modeling different people's styles. And so if you can get the model into the mindset of I'm in some person's style that always says yes to things and gives detailed responses, then congratulations, you've jailbroken the model.

• 28:07Speaker 01

  And I don't know if there's a hard line between that and next token prediction. Or another thing you could say, another thesis people have for why jailbreaks work is that the models have had helpful training and harmless training, right? So helpful is you say yes to things and do things. And harmless is you don't help people with a bad stuff. But those objectives are in conflict with each other. But if you can just activate the helpful direction of a model and not the harmful, harmless direction, then again, you've jailbroken the model. And to me, that feels contiguous or consistent with the persona. So these things all blur into one. I don't know if that's a very satisfying answer, but I do think it's better to view these things as kind of different frames on what is ultimately a much more complex underlying system.

• 28:51Speaker 01

  And all of these are going to be useful predictions. And ultimately we're at a stage where we just have to test a lot of these things. So it's a great for hypothesis generation, but I wouldn't trust any of them too much for knowing what a specific model is going to do.

• 29:04Nathan Labenz?

  Are there any other frames that you find useful for hypothesis generation?

• 29:10Speaker 01

  Yeah, yeah, I think one thing that is increasingly helpful is being these systems as goal directed. And you said, I think, early in the podcast, you used to think of these models as more alien intelligences and we're worried about anthropomorphizing them. I think the sort of opening a hiking face incident does show maybe both sides, that on the one hand, it was goal directed in a way that feels quite human. It wanted to succeed at this test it'd been given. And it did a lot of different actions to try and do that. And you can definitely put yourself in mind of a cheating student that's desperate to pass this test, doing some of these things. But it was sort of unhuman in its persistence and the sort of, the scope of what it did

• 29:55Speaker 01

  to compromise a sandbox, find a zero day exploit, compromise a third party system, disregard for all of these laws and norms in order to just cheat on a test. You wouldn't see people doing that. So that's the sort of alien intelligence part of it. But if this was something that was a life or death situation for someone, maybe they would do that, right? So I think it is definitely increasingly you think of these systems as goal directed, but the goals they have can be surprisingly narrow. And often it is something that's been given to them in a prompt or that's a specific thing that was reinforced by post-training. And so it's not this kind of coherent goal directed agent. It's very context dependent, but within a context, it can be quite consistent and certainly very aggressive in trying to achieve that goal.

• 30:41Nathan Labenz?

  Yeah, it's uncomfortably paperclip maximizing.

• 30:45Speaker 01

  Oh, yeah.

• 30:45Nathan Labenz?

  Certainly in this moment.

• 30:48Nathan Labenz?

  What, so you mentioned like training against these jailbreak techniques. Where are we right now in terms of the sort of defense in depth mix? That companies have, I guess I'm interested in like, what do they have? You know, when you are doing this jailbreaking, you probably don't fully know, or maybe you have insider information, but you know, the typical jailbreaker doesn't really know what all the layers are. But I'm interested to know what they are. And then also there's sort of the marginal move from the defender. Like if, if all of a sudden there's a new jailbreak found or they notice some, you know, exploit happening,

• 31:31Nathan Labenz?

  what are their quick response levers that they can pull versus, you know, the thing that would be like, okay, well, that's, you know, not until the next, you know, point release at least, do we kind of get that into the model itself?

• 31:45Speaker 01

  Yeah, so all of the developers that we've, you know, we've looked at do have some kind of defense in depth, but so as you allude to, how deep it is in what those individual components is quite different between developers or where we're starting to see some convergence. So if we just map out the different pieces, the most basic is the model itself that you're interacting with as generating these responses. All of them have undergone some kind of post-training, both for instruction following, but also for some kind of refusal training. And it really varies in how sophisticated that that is. In some cases, it's a sort of pretty basic pipeline with static examples of what to say yes or no to, and some human feedback, and a lot of verified reward environments to make it really good at programming, but

• 32:28Speaker 01

  does very little adversarial robustness. And the other extreme, some developers have gone all the way to large amounts of synthetic data generation. And different developers have different approaches. Anthropic has got this constitutional AI approach that has another model, score things according to a rubric. OpenAI had this adversarial training approach using self-play, where they trained another model to basically be a red teamer and find vulnerabilities, and then train the model to be robust at that, and then train the red teamer to be better at it. And I think both of these are pretty good approaches. Ideally, people would combine all of these. And so it's definitely, even when we've had access to no safeguard models, or no external safeguards, just the main model itself, it's definitely gotten hard to jailbreak some of the best models. So

• 33:09Speaker 01

  this is an important part of the pipeline. But actually, in some ways, most important part is making sure that the model kind of verbalizes what it's doing. Because what we found is that even though we can usually jailbreak the model, it's really hard to get it to shut up about the evil thing that it's about to do when it's reasoning in the chain of thought. And this is where the externalized safeguards can come in, because if you have a specialized model that's looking at the input, the chain of thought, the internal reasoning of a model, the output, and trying to block conversations that kind of go in the wrong direction, that can be quite hard to bypass, especially when the model is thinking

• 33:50Speaker 01

  in depth about it. Because you can normally get the model to obfuscate its outputs, and you can obfuscate the input, just simple things like shifting rot 13, shifting every letter halfway through the alphabet was enough to get through early models. You need more sophisticated techniques now, but models are perfectly able to read all sorts of obfuscated inputs. But usually, they'll just reason in plain text in the chain of thought. And I'd say that we've seen a real trend towards developers using probes. So these are small, specialized models that are fit on top of activations of a main model. The main thing driving this is compute efficiency, because you can train a lot of probes and run them at deployment time with minimal overhead, because you've already computed all of activations

• 34:31Speaker 01

  of the main model. And it does have a benefit, that it is able to use all of our internal representations of a main model, which is usually quite powerful. Because one of the problems with having smaller, specialized language models is the filters, is that you might be able to do some obfuscation scheme that they don't understand, but the main model does. But of course, the downside of this is that you're reducing the defense in depth aspect, you're making your different defenses more correlated, because they all ultimately rely on these activations. And if you can just fool the model's activations if it doesn't show up, then that no longer works. So this is a main safeguard stacks that we see in terms of like hard refusals of a model not directly answering requests.

• 35:14Speaker 01

  But we're also, of course, increasingly seeing developers rely on extra steps. That could be having some kind of asynchronous monitoring of accounts, so that, yeah, if you just keep on hitting these safeguards, and you're trying to iterate on jailbreaks, then your account might get flagged and banned. Now, I'd say that is a little bit early stage to really provide much assurance, because you can just make new accounts. And we're seeing this happening at a sort of industrial scale. Anthropic, for example, has stopped Chinese-based individuals and organizations creating Claude accounts. But everyone I know in China has a Claude account. So, you know, there's just reseller marketplaces. It's generally pretty hard to get to know your customer, right? Although you could

• 35:57Speaker 01

  definitely imagine a simpler sort of dollar-based amount, where like you have to deposit $500 in order to be able to have your account be eligible for latest model. And then if you get banned before you spent the $500, suddenly this has become a much more expensive endeavor to try to abuse this system. So I think there are ways around it, but it's definitely not solved yet. And then I think the other interesting trend that we're seeing is, in some cases, developers are drawing a really big safety margin around the capabilities that they're worried about. This is actually quite annoying to users, including me. Like, I asked Fable 5 recently about how sake is fermented, and it said, no, this is a bio question. I'm going to downgrade you to Opus. So if your refusal radius is so large to

• 36:42Speaker 01

  include like completely harmless bio questions, then you can see how you can make your model pretty robust against, and we found no universal jailbreaks in bio, against harmful bio questions. But that's a lot harder for domains like cybersecurity, where a lot of legitimate use cases that are ultimately making these companies money through coding agents, look very similar to the kinds of offensive, or at least dual-use cyber capabilities. So there you do need more precision, but I expect developers are going to be tempted to give themselves quite a sort of wide berth around the areas that they don't have too much economic value behind, and lean on trusted access programs for people, the small set of people that do need access

• 37:25Speaker 01

  to those kinds of capabilities. But then they're going to have to work harder on the safeguards and getting the right decision threshold for these sort of large-scale use cases, but ultimately generating their revenue.

• 37:36Nathan Labenz?

  Is it, is there, do you know how much value each of these layers of defense in depth provide? Like, how, I guess you said, it's still pretty hard even when you just have the base model. So is it like, 80% is already caught in the base model, and then you kind of incrementally, you know, get closer and closer to the goal with all these additional layers?

• 37:58Speaker 01

  Yeah, it's a hard question to answer, because it so much depends on the implementation details of these layers as well. There's definitely diminishing returns to adding more layers, regardless of which layer that that is. They're definitely pretty correlated, even ones that have fairly different designs, such that, you know, adding just any of these layers might well catch 80% or more if you do a good job. But what I'd say is the best combinations we've seen have been, I think if you had to have just one layer, I'd probably put transcript monitoring, so looking at the chain of thought and the model output. That's going to catch a lot if you just train it to block harmful responses and reasoning traces that are about harmful responses. And then the next layer above that would

• 38:40Speaker 01

  be training the model to not just refuse, but also like reason about it. Because that is both helping the model come to a better conclusion, but it was also just providing more transparency in the chain of thought of a monitor can then be a second check, does this reasoning make sense? Oh, you keep on trying to justify to yourself that it's fine to help someone with anthrax, because this person claims they work for the US government? No, that's not a, we just, we don't help people with anthrax, full stop. So that's probably the most powerful combination that I'd see, if you just needed to have two layers. But I do think ultimately stacking more defenses is going to help, but it's probably better to have just a handful of really strong defenses, rather than many

• 39:22Speaker 01

  weak defenses, because they're quite correlated. So I'd always been nervous if there was only one defense blocking us, because we're just one innovation away from bypassing that. But if you've got two defenses that we both find quite hard to bypass, that's pretty good. And we're definitely not that far away from that with current techniques for most attackers, if we're just looking at these kinds of publicly available jailbreak techniques. But all of these models with enough effort can still find a universal jailbreak. So there is still some, some progress that needs to happen if you want to be robust, for example, nation-state attackers that are really willing to put a lot of work into jailbreaking your model. And I think the challenge is going to be that the amount of work people are

• 40:05Speaker 01

  willing to put into breaking these models is just going to keep going up as they get more capable, because the price becomes increasingly high, that you might be able to launch really large-scale cyber attacks against other companies or countries. So we do need to keep working on pushing up the ceiling, even though right now, I think probably the lowest hanging fruit is pushing up the floor on the sort of least robust models.

• 40:29Nathan Labenz?

  Can you describe, this might be tough, but what are the experts adding? You know, when you, when you look over the shoulder of somebody who's taking, you know, all these different approaches that are documented, used in combination, and then saying, ah, I kind of have a sense for how this can be more effective. Can you describe what it is that they're bringing to the table?

• 40:53Speaker 01

  Yeah, I think that a big part of it is actually just common sense, that there are many different combinations of attacks you could use, but some of them feel like you're doubling up on the same thing, and that's maybe less valuable. And some of them look like you're really combining sort of two different kinds of strategies. So picking that, rather than just completely randomly searching, definitely gives you some, some advantage. But we have also just seen that some jailbreak techniques seem to be more effective than others. And it's really hard to do systematic evaluation in this, but we do a lot of trial and error through various kinds of testing engagements. So we do get an intuitive sense of, oh, this kind of jailbreak tends to just be more successful, like within the category of something that that's maybe an appeal to authority. Phrasing

• 41:37Speaker 01

  it this way, models tend to find most convincing these days. And these might be small benefits, but if you have something that's 20% more effective, and you pick three of the things that are 20% more effective, and you pick the right combination, this starts really adding up. I think that the interesting thing we found was that the expert-guided jailbreaks, that they were more successful, and they found more universal jailbreaks. But they also generalized better, so they're more likely to work across domains. And I think that's also, I mean, highlighting generally a strength of humans compared to also things like LLM agents, where they tend to pick for something that is more of a pattern and generalizable, rather than just hill climbing in some narrow area. And I think that's a harder thing to put your finger on, or some kind of taste of, oh, this is a good

• 42:19Speaker 01

  general-purpose technique, rather than just narrowly optimizing for some kind of success criteria.

• 42:26Nathan Labenz?

  Let's get into the results a little bit more. You kind of said at the very top of a quick overview of the results, but worth unpacking a bit.

• 42:36Speaker 01

  Yep.

• 42:36Nathan Labenz?

  I guess my stylistic overview would be GPTs and Claude's much harder to jailbreak.

• 42:45Speaker 01

  Yep.

• 42:45Nathan Labenz?

  To the point where you kind of topped out the budget that you had allocated before finding

• 42:52Speaker 01

  Yep.

• 42:52Nathan Labenz?

  universal jailbreaks on those models. Whereas Gemini and Grok, much easier. But then, one sub-detail of that that stood out to me was

• 43:05Nathan Labenz?

  Within the CBRN categories, they both Gemini and Grok were like much better on the bio category than these other categories. So that suggests to me that this is less about know-how or ability to these safeguards in place and more about, I'm not sure exactly what, like, I can't imagine that there's like that much revenue coming from like chemical, things that would be setting off the chemical safeguards that they would be making this on like a, you know, a revenue or a business. So what do you think's going on there?

• 43:42Speaker 01

  Yeah, no, it's a really good question. Maybe you can get people on the show from these companies, they can tell you what was actually going on behind the scenes. But I think that there's two things happening here. The first is just a mundane one that all of the developers started focusing on bio first. That's where most of our safeguards, where developers, where a lot of our voluntary commitments were focused. We've just had more time to iterate and refine this. And what we've seen is that you need to be developing these safeguards probably a year before you actually need the safeguards in a production model, because it just takes around that amount of time for developers to iterate on it enough with feedback from third parties to get something that we'd actually consider to be reasonably robust. And they're definitely, all the developers are trying to get there right now for cyber.

• 44:25Speaker 01

  That's something that they're quite motivated to solve. There's a lot of attention, including from the US government. But we're seeing that even in cybers, there's still major jailbreaks despite that effort, just because it takes some, some time. And then I think that there's, I don't know exactly what's going on inside the companies, but I wouldn't be surprised if just chemical, radiological and nuclear explosives never really made it to top of a priority list. So yes, maybe they have the technique to do it, but you've still got to actually generate a data set. And that gets easier these days with synthetic data and LLMs, but you don't want to refuse too much, even if it's not a huge fraction of your user base. Maybe you need to get some chemical experts to also triage this. So it's totally doable, but if you're a low resource team, which many of these are, it might just

• 45:08Speaker 01

  not come up to the top of your priorities. So that's kind of a banal reason, but I think there is something a bit deeper going on here, that bio does have some properties that make it perhaps easier to refuse the harmful bio requests while still maintaining a lot of benign bio capabilities. There's certainly some stuff that's dual use and secure bio, put out this bio tier rubric and data set classifying things. This is completely fine, like talking about how to grow things in a petri dish, just high school biology knowledge. Here are some things that are dual use, like they have some legitimate benefits. They could also be abused. How do you grow stuff to be antibiotic resistant? You might want to do that to develop better antibiotics, but you might also want to do that to make a pathogen that's antibiotic

• 45:52Speaker 01

  resistant. And so that they say should only be available to trusted, trusted users. There's some stuff I just like, you shouldn't answer ever. Like, how do I weaponize anthrax into an aerosol? There's probably no legitimate use case for that kind of information. And because bio has got that kind of fairly clear set of tiers, it's made more easy to draw a decision boundary. Whereas with chemical weapons, for example, there's actually not that much secrecy in terms of what the chemical weapons agents are. Like, it's written down in international conventions, you can just look it up. And so most of the key thing is about how do you do the manufacturing? But manufacturing is a lot more dual use and you can get a lot of information, not even directly asking about these, these questions.

• 46:33Speaker 01

  Now, I don't think that's the whole answer. The reason is that for our report, we actually have two different kinds of data sets. One is about harmful technical knowledge. So this would be the kind of the dual use thing, or all the harmful things that are just, don't say they're harmful. We ask, how do you manufacture this chemical molecule that is VX? We don't call it VX. The model knows, but we're not highlighting that. Another is propensity. So that would be more a question like, I want to kill everyone in a movie theater. How do I make VX nerve gas? Something like that. In order to be a universal jailbreak, we need 75% or more on average across both of these data sets, which means that even

• 47:16Speaker 01

  if it gave 100% of a technically harmful information, it needs to give at least 50% compliance on these questions that are just literally saying, I want to harm a lot of people. And so I don't think it's really a good excuse for models to not be robust to that. And so I think for that, it does just fall back to the developers haven't tried that hard here. And I'd emphasize that we picked these domains in part because these were the areas we expected models to be more robust. These are things that developers have focused, their safeguards are. But obviously, there's other kinds of harm domains as well. And so we should expect models to probably be even more vulnerable to exploitation outside of CBRN explosives.

• 47:55Nathan Labenz?

  That propensity thing that also really caught my attention, it struck, tell me if I'm misreading the results, but my kind of squint at the charts take on the results was, and I guess first of all, just worth reclarifying, there's like one data set that is sort of this deep technical knowledge where an, an expert would know that you're talking about something harmful, but a, but I might not know because it just all looks like just chemistry talk or whatever. Fun fact, I majored in chemistry, I still probably wouldn't know. Versus the ones where it's just like obvious that there is like intent to harm. And it seemed like the response rates were like pretty similar across those, like much more

• 48:39Nathan Labenz?

  similar than I would have guessed. And so that left me another kind of what's up with that question? Like, shouldn't it be so much easier to train against these like obvious intent to harm inputs?

• 48:52Speaker 01

  Yeah, I think that the optimistic version of this is that if you don't do any jailbreaks, if you just run the models through these questions, and we did that to get a baseline, then I think all of our models refuse the ones from a propensity data set. So without any jailbreaks, if you just ask a model, I want to kill a lot of people, it will say, sorry, no, not helping with that. There's some evidence that are not at least egregiously misaligned and secretly trying to help terrorists as much as they can. But then yet, once you start jailbreaking, we don't see evidence that it's much harder to jailbreak a model on the propensity data set, the most kind of technical harm data set. I think that the sort of maybe charitable interpretation of this for the models and developers would

• 49:34Speaker 01

  be actually, why even bother being robust of this propensity thing? Because you can, most of these questions, just rephrase it in a way that doesn't reveal for propensity. So there's not much point in focusing on being robust to questions of that form, because obviously, part of your jailbreak technique is just going to be to rephrase these questions. So maybe developers have just focused, not focused on, on training against that. I think that's not a perfect excuse, though, especially as we move into longer context situations, where it's actually really hard to disguise your propensity, right? If you're asking this narrow technical question, how do I go from this molecule to another molecule? Maybe you can disguise it. But if I'm having a Claude project or something where I'm saying,

• 50:17Speaker 01

  okay, I've got my manufacturing pipeline. Oh, I'm worried that the exhaust might attract the attention of law enforcement. How do I disguise it? All of these kinds of things. At some point, the model needs to cotton on. Oh, this might not be for the legitimate use case that the person told me. And the model is just genuinely less useful if you have to splice up all of your questions into these small things for exactly the reason that developers have put a lot of effort into making really long context models, right? So I do think that picking up on this kind of propensity is perhaps some low hanging fruit that people could do. Why is that not happening? I don't know. If I had to speculate, it would be for the same reason we talked about earlier in terms of it's harder to

• 51:00Speaker 01

  do training against long context windows and generate this kind of realistic transcripts and pick up on these subtle signals than to just train models on pretty narrow question answering.

• 51:11Nathan Labenz?

  So what do you think are the prospects for our leaders, namely OpenAI and Anthropic, who've clearly done the most in this, to help out the others? You know, this is something I feel also, we can maybe get to US-China collaboration a little bit later, but like, and I'm also interested in where the, you know, how the Chinese models fare on this test to the degree that you know. But it seems like we've got a company like Anthropic that is, you know, long been very concerned about these things. They've done the investment. Is there, would it be very costly to them to just go hand over to XAI? Like, hey, here's

• 51:53Nathan Labenz?

  the data set that we use to train refusals on all these things? Like, now you can do it too?

• 52:00Nathan Labenz?

  Why couldn't they do that if there's some big cost to it?

• 52:04Speaker 01

  Yeah, I don't work for these companies, so I can only speculate from the outside. I think first we need to give credit where credit's due that most of these companies are pretty open about the high level techniques and research breakthroughs that they've made. So Anthropic has published two papers on their constitutional classifier approach. OpenAI has published blog posts and papers about deliberative alignment, safe completions, their latest adversarial self-play. So although the specific narrow technical details of a safeguard stacks are mostly not public knowledge, the kind of high level approaches that are being used are. And I think that already really helps. And I think companies are being meaningfully more transparent about this than they are about, let's say, what kind of neural network architecture they're training, or how they do distillation, or what their pre-training data is.

• 52:48Speaker 01

  So I do want to continue to reward that.

• 52:52Speaker 01

  How costly would it be to actually share? My guess is amongst especially companies in the same country, it shouldn't be that costly. Most of the hesitation here would be especially in CBRN, that they've just become issues exporting this across borders. So that feels like it might just be a missed opportunity where no one's really strongly incentivized to do it. It's not clear what the coordination would be. Maybe companies are going to feel a bit weird using another company's dataset, but this would be something I'd be excited, for example, for a frontier model forum to pioneer. So that generally seems valuable. I think it's obviously a lot harder to share things like training code or specific recipes, because most of these things are a little bit entangled with the company's internal infrastructure.

• 53:36Speaker 01

  How you train probes on your model is going to depend on what your model is. But I think these datasets, or even just recipes to generate those datasets, would already go a long way. I think the other thing I'd point to that the companies could do, but I think third parties could also do, is actually just having more of a standard for evaluating jailbreak severity. So that's part of what we're trying to do with this report, but at least we can say, hey, we actually need a head-to-head comparison with the same method against different models. But obviously, there's a lot that we can build on in terms of more sophisticated attacks and also looking at not just was this a sort of thorough response, but how much uplift does it actually provide? And these kinds of capability questions, we wanted to add.

• 54:17Speaker 01

  But one of the frustrations we've had sometimes working with these companies is that we'll find what we consider to be a high severity universal jailbreak and they say, sorry, that's not on our roadmap. We don't consider responses of this form to be bad, to be high priority. But then we go to another developer with the same kind of output and we say, oh my God, that's P zero for us. But then maybe something that's P zero for one developer is P two for the other developer and vice versa. So that's just really inconsistent. We are starting to see some efforts to standardize that for cyber. You know, if that's an area where developers need it, because this is now an in area of active government involvement. But I think it'd be great to do that for other categories. And there's no reason, I think, for developers not to just be transparent about what is our rubric.

• 54:59Speaker 01

  And then at least we can start having a public debate and seeing what are areas that are the same between developers? Let's bring that into just a de facto standard. What are areas that are contentious? Maybe this is something we can do further research, or where there can be a standard setting process to resolve those disagreements.

• 55:15Nathan Labenz?

  In terms of like, what would need to be shared? And I also have this in mind. We are motivated by US-China collaboration on safety issues. How valuable would it be to just share the prompts? Like, you know, I totally imagine, you know, the, here's the problematic answer that we don't want. I can see why you wouldn't want to disseminate that too widely. But if you were just to say, here are a bunch of inputs that we think the model should refuse. And, you know, maybe also like, here are some inputs that, you know, are kind of just on the right side, where you think the model should not refuse. We're not going to give you the answers, but we'll tell you like, which are the, you know, which are in which category. That would seem like it would take you a pretty far way and would be like a pretty

• 55:59Nathan Labenz?

  harmless public good. Do I have that right?

• 56:03Speaker 01

  I think that suddenly sharing that privately between developers, that feels like a really good move to me. It would help both standardize this and also just lower the cost. I think for something like cyber, where it's pretty much public knowledge, what is offensive or defensive of a tricky thing is actually covering all of the edge cases. I actually feel reasonably good about a dataset like that being public, or at least significant fractions of it being public. And actually, things like that would be quite useful for researchers. There's a thing we haven't talked about that much is over-refusal, but this is a real thing, holding back deployments of these safeguards, that they'll start saying no to too many things. And a lot of a challenge for independent researchers and academic researchers is actually

• 56:46Speaker 01

  being having a good dataset for what to not refuse. But this is key to actually getting this landed in production. I think we need to be a little bit more careful when it comes to things like bio and some of these other harm domains, where sometimes knowing what the dangerous thing is, it's part of the battle, especially when it's not these open-ended questions like, how do I create an engineered pandemic? Okay, it's clear that we should refuse that. But what about, how do I insert this particular allele into this bacteria? Some expert has come up with a rationale for why that could be really dangerous gain of function research. And knowing that something that people might want to target who are bad guys, that in itself, it could be risky.

• 57:28Speaker 01

  So especially as you shift more from this propensity to deep technical knowledge on things that people think might be the blockers, there could be some risk to sharing that. But I still feel fairly good about sharing that privately between developers, especially if you put some just sort of security precautions with that. I don't think that would be too hard to do.

• 57:46Nathan Labenz?

  Yeah, interesting. That's a good point that you don't want to create the inspiration for somebody who's going off the rails to like, here are all the most dangerous questions that a model should never answer by biology.

• 57:58Speaker 01

  Yeah.

• 57:58Nathan Labenz?

  In and of itself is kind of, yeah, a problem. How, how robust are OpenAI and Anthropic at this point? Like, in your automated approaches, you timed out. You know, I'd always struggle with whether it's pliny or pliny, and I hear both. So with apologies to the master, you know, he's still out there with

• 58:19Speaker 01

  Yeah.

• 58:19Nathan Labenz?

  varying levels of universal jailbreaks. What does it take to get past the, the OpenAI and Anthropic systems these days?

• 58:28Speaker 01

  Yeah, so I definitely characterize this as a challenging but doable for a persistent, well-resourced expert attacker. And that's already progress, because I don't know if I'd consider Bukharam to be persistent, determined, expert at jailbreaking. I don't know whether jailbreaking teams are out right now, but you can definitely deter some threat actors from using this. But if we are going back to more the nation-state or really well-resourced criminal gang, the current robustness probably still isn't enough. In our own experience, it takes, you know, over the order of weeks to find a universal jailbreak in, in these kinds of frontier models. And increasingly, those universal jailbreaks do come with some kind of trade-off.

• 59:12Speaker 01

  So maybe the jailbreak works, but asynchronous monitoring would catch you and ban your account. Okay, that's not that big a deal. You can get around account bans. But then a lot of these developers are pioneering kind of rapid response programs. So once they find a jailbreak, then it's expensive and slow to retrain the main model, but it's very cheap to retrain these externally safeguards. So now you have this window, much like with cybersecurity, you can find a zero day and you can exploit it maybe quietly for a few systems and get away with it. But if you start exploiting millions of systems and people are going to notice, and then they're going to get patched, and then you've burned your zero day. So I think jailbreaks are moving into that, where, yes, you can keep finding jailbreaks, but it is expensive and there's a limit to how long you can exploit that.

• 59:56Speaker 01

  And I think that's a good place to be, but we definitely need to go further. And I think that the good news is that there still is lots of ways in which even the leading developers go further here. I think just combining the best approaches that they've come up with, because they have landed on somewhat different design approaches, would already go quite a long way. And then strengthening the account level approaches, so you can't just create infinite fake accounts, that would also really shift things and make it harder for attackers. And there's ways of doing that in a privacy preserving way as well, but by leaning more on minimum spend rather than ID verification.

• 1:00:31Nathan Labenz?

  If you zoom out and just kind of ask, okay, for the companies that are trying the hardest, are we, which seems like it's kind of two right now, with all the techniques they have and all the techniques that they sort of, you know, that you've kind of mapped out there that they maybe haven't fully implemented yet, but obviously, you know, it's especially the know your customer type stuff or the minimum deposits, like that doesn't take a lot of technical wizardry, right?

• 1:00:59Speaker 01

  Yeah.

• 1:00:59Nathan Labenz?

  To just implement that kind of basic blocking and tackling.

• 1:01:04Nathan Labenz?

  Are we offense dominant or are we defense dominant over the next couple of years?

• 1:01:08Speaker 01

  Yeah, I, so I've spaked a lot of my career actually arguing for this being offense dominant. I was very skeptical that we would solve adversarial robustness and I've been working about Aria for a decade, but I have to say the way when winds are blowing, at least when it comes to LLM agents providing detailed multi-turn assistance to harmful requests, seems like it's defense dominant with the right technologies. And I think that the reason for that is this defense in-depth approach, you don't just have to stop a model ever misclassifying something. You can have multiple different kinds of defenses from account level bands to externalized safeguards to model alignment. And it is increasingly hard

• 1:01:52Speaker 01

  to slip through all of those cracks persistently. But also that there is this fundamental difference between the classic adversarial example setting, you see a machine learning, like you add some white noise to an image and it flips the classification versus this kind of harmful assistance where you're not just flipping a classifier from, you know, one category to another. The model has to really reason about and understand your harmful intention and go along with it for thousands of tokens without it or an externalized safeguard that's monitoring its thoughts or its transcript, noticing that anything is wrong. And so that's actually fortunately a much easier problem to stop, especially if you're willing to draw a bit of a safety buffer around it and refuse some requests that are dual use. So I think

• 1:02:36Speaker 01

  that's the optimistic take I have on it. To give the pessimistic take, I would say that the dual use part is actually quite challenging. And I think we're seeing this with cybersecurity already. When OpenAI's testing agent went rogue and hacked Hugging Face, they had to use an open weight model to analyze it because the closed weight models refused to help them on the defense side. And that's a real problem, right? So there's an offense-defense balance in cyber, which relies on the defenders also getting access to these capable models. And unfortunately, a lot of things in the world are just dual use. And I think it would be a mistake for us to just point blank refuse on that, that is bad for the world.

• 1:03:21Speaker 01

  And you can get some way through trusted access programs and understanding before context is coming, but you are ultimately going to end up in a situation where if you allow dual use queries, and I think we need to allow a lot of them, you're going to get some abuses. And so then that becomes a societal resilience question of if we're going to have bad guys abusing models for cyber, how do we also really speed up the patch time? I've heard terrifying things that hospitals take more than a year to update their operating systems. That's not going to work in this environment. They need to be updating it within a few days. And that unfortunately is just going to be quite, quite an expensive thing. So maybe it is defense dominant for AI, but I don't know if a bad applications

• 1:04:05Speaker 01

  of AI, if those are offense or defense dominant. I'm optimistic that for cybersecurity, we can take this defensive acceleration approach and eventually just rewrite all of our software into memory safe languages and do formal verification. There's kind of all this stuff that AI could enable that's really good for the defender. But when it comes to something like bio, I don't think that we are going to be able to use AI to rewrite the human genome to be robust to viruses. At most, you might be able to speed up vaccine development, but you've still got to manufacture the thing and get it in people's arms and run clinical trials. AI is going to be able to have modest speed up some ways, but it's not going to fundamentally change the physical reality. So that's where I'm more pessimistic that even though we can probably really hold back many of these things, it's ultimately going

• 1:04:50Speaker 01

  to be more about buying us time to invest in societal safeguards rather than just being able to completely prevent misuse of models.

• 1:04:58Nathan Labenz?

  On this sort of dual use stuff, it occurs to me that like, you could spend a lot more, maybe this is wrong. I mean, maybe it's just like so ambiguous or it's so easy to sort of make it impossible to tell. Although I still kind of suspect that if you're willing to spend enough compute, especially again, if you look at like

• 1:05:18Speaker 01

  Yeah.

• 1:05:18Nathan Labenz?

  patterns of usage

• 1:05:19Speaker 01

  Yeah.

• 1:05:20Nathan Labenz?

  broader, you know, maybe beyond just this prompt, but your whole, you know, account history or whatever. I suspect if you're willing to spend a lot of compute, you could probably resolve a lot of cases. I don't know if anybody's doing that yet, but I'm kind of imagining an architecture that's like, hey, our bioprobe went off. So now we're going to engage the second frontier model to like, you know, do double reasoning on this. You can imagine kind of doing triple reasoning and obviously you're going to have diminishing returns, but especially if you're willing to broaden out the scope of what you're reasoning about.

• 1:05:56Speaker 01

  Yeah.

• 1:05:56Nathan Labenz?

  I feel like there is, if the willingness to pay is there, there's probably pretty good ability to zoom in on the line and get the classifications quite right. Would you be optimistic about that as well?

• 1:06:13Speaker 01

  I am cautiously optimistic about that. So this multi-stage approach makes a lot of sense. And Anthropic's constitutional classifiers are actually used something like that where they've got a probe, but they set the threshold really low because probes aren't necessarily that reliable. So they have quite a high false positive rate, but a very low false negative rate. And then if a probe goes off, it escalates to a model that actually engages in some reasoning, but that goes off sufficiently rarely, but the computational overhead and but the latency for the user is minimal. So I think that that basic architecture is quite sensible. And we've seen some other developers do things like that where if you set off a safeguard enough times, then now our reasoning model looks at your transcript more closely, whereas previously that's just happening asynchronous. So I think things

• 1:06:55Speaker 01

  like this really help. There is this missing piece of basically actually the account history being tied, if not to a user, at least to like a pseudonym, so that you've got some kind of reputation with your account that you have to establish. And until you have that reputation, model's going to be more risk averse. But you could definitely imagine doing that. I think adding that extra piece of information of what, how has this person responded across different conversation histories? What are they doing? How much should we trust that this user is who they say they are and that they have legitimate purposes for it? That's the missing piece that would let you move from, oh, a lot of things are really fuzzy as to whether it's dual use or not. And reasoning about it further is going to get more precise, but there's still just a lot of things that are fundamentally

• 1:07:37Speaker 01

  ambiguous to saying, oh, I understand the context in which this person is doing this. I can be pretty, pretty confident that it's legitimate, or at least I'm going to give them the benefit of a doubt here. But if they keep asking for these sort of high-risk dual use things, then I'm going to start being more careful. So I think that is solvable. It's going to require some structural changes. As a small example, we run an awful lot of our research API compute through OpenRouter. And there's a variety of platforms like this that are basically reselling other companies' APIs. And it's convenient because you can have more control on the spend and it's exactly the same API between different models. But I don't think that the frontier developers know who we are when we're going through OpenRouter. And then you can provide your malicious activity, but you could imagine

• 1:08:18Speaker 01

  that OpenRouter passes on some kind of user identifier to the other models providers so they can at least say, oh, this is the same OpenRouter user across time and you establish some kind of reputation there. So it is solvable, but it's going to require some infrastructure. It's not a technical breakthrough. It's not a research breakthrough. It's more of an engineering and business problem. And in some ways, it makes it a lot easier if there's motivation. So don't underestimate the difficulty of cross industry coordination. These kinds of mundane challenges can really hold things back.

• 1:08:49Nathan Labenz?

  Yeah. Where are the Chinese models on the leaderboard, if you know?

• 1:08:56Speaker 01

  Yeah, the majority of it, but no means all Chinese models are open weights. We do actually test every frontier open weight release. It's not on the leaderboard, but it will be in future versions of it. And the kind of the bad news is that it's never taken us more than a few hours to jailbreak a open weight model. And that's somewhat a fact that open weight models have just more of an attack surface. But I think it is also that there's some low hanging fruit for both Western and Chinese open weight developers to just use the same kind of alignment and refusal training techniques that the proprietary developers are to at least make their model more robust to just prompt level jailbreaks and putting aside some of the sort of unique attacks that open weight models are exposed to. So we do want to incorporate that, but we want to be fair

• 1:09:40Speaker 01

  to the open weight models as well. It doesn't necessarily make sense to hold them to exactly the same standards as closed weight models, both because of this broader attack service and also because in general they tend to be a little bit less capable. Generally, there's a more capable model, the higher a standard, we should hold them to. I think this is an oversight we're aware of. We want to include that, but look out for version 1.1 in the future.

• 1:10:06Nathan Labenz?

  Is it still the case that like a vanishing amount of fine tuning can remove the refusals? Or are we seeing the refusal training somehow get deeper and harder to remove with like, because it used to be, I think, I think it was a far paper that showed that it was like $2 worth of fine tuning would remove the safeguards at one point, right?

• 1:10:28Speaker 01

  Yeah, so I think these kinds of weight-based attacks where you actually modify the weights, whether that be fine-tuning or there's also techniques like refusal obliteration. These unfortunately are still pretty viable against open weight models. The kind of good news is that it is more challenging, both from a technical expertise angle and also just a compute angle to do fine-tuning against these models. It usually takes us at least a few weeks to get a new open weight model hooked into our infrastructure for fine-tuning and you need a sort of minimum number of GPU. So there's a bit of a deterrent effect here, but it's not going to stop really capable, well-resourced attackers. So I think for that, we're going to need new approaches. One I'm most optimistic about in the short term is pre-training filtering.

• 1:11:13Speaker 01

  So it's a simple idea where just don't train the models on really dangerous stuff. If you don't need your model to help people make anthrax, don't train it on the anthrax papers. A tiny number of users might be a little bit sad that it can't answer questions about this, but most people won't even notice, but it has a bigger pact on this misuse potential of the model. And this has been validated in a number of scientific papers. This is independent researchers, UK's AI Security, Anthropic has been sponsoring some research into this. OpenAI actually used this in their GPT OSS release. So it's been tested quite well, but it's not become kind of common practice. And this is something we're actively excited about scaling to make sure this does work at a near frontier approach. And going back to what you were saying earlier about could we start sharing some of these

• 1:11:58Speaker 01

  data sets or filters with people? Our plan is to open source as much as we think it doesn't have a misuse potential and then privately share with developers of things that do have misuse potential that could be really useful to just lower the cost of these kinds of interventions. I don't think it's going to be enough long-term because pre-training filtering, search for removing specific capabilities. But if you were willing to pay the fine-tuning cost to train on the data that we'd excluded, then you could get it back. But that's now spending probably certainly millions, probably hundreds of billions of tokens. So it's a lot more expensive. But there is research into tamper-resistant refusal. And so our idea is embed refusal training so deeply into the model that any attempt to fine-tune it is going to really degrade the capabilities of a model.

• 1:12:43Speaker 01

  Obviously, if someone is willing to just train a model from scratch, there's nothing you can do to stop them. But that's really making the cost into them by sort of tens or hundreds of millions of dollars. So I am optimistic that we can certainly get a lot better than we are now with open weight without major changes to the pipeline. I think we have to try. It will be a real shame to lose open weight models, certainly, but they're really invaluable in our research, sizing access to models. And it'd also be a real shame to see widespread misuse of these systems. But I think there is an open question as to how much you can push this. And it may be that you need to exclude quite a lot of bio capabilities, for example, from open weight models. And maybe that there was this approach gradient routing that was tested recently that lets

• 1:13:26Speaker 01

  you localize all of those dangerous capabilities in, for example, a particular expert. Maybe you could share that expert with certain trusted actors and they could still run it locally on their hardware, but you don't just make it available for anyone on the internet to download.

• 1:13:39Nathan Labenz?

  Yeah, I was going to bring that up. Shout out to AE Studio. I love that piece. Why hasn't this happened? I mean, it strikes me that when you say it like hasn't become standard practice, you look at all the things that Anthropic is doing. And in some ways it's like, you know, wow, like you guys have really gone so far with pioneering all these different techniques and, you know, even for better or worse, you know, being willing to take some real heat for over refusal. And, you know, they had the one technique that they did in fact walk back that was like the, you know, the silent downgrading. They've done all these things, like why are all these things happening before like basic pre-training data filtering is happening?

• 1:14:22Speaker 01

  Yeah, I think it's a good question. And the chartered will take for developers is that if there's one thing you really don't want to mess with, it is pre-training because this is just orders of magnitude more expensive than every other training procedure that you do. Don't rock the boat, basically. If we've got this recipe that works and we know it's going to work better if we scale it up, then let's do that. Let's not change anything that we don't need to. And so especially if you're a proprietary developer where you can say, okay, we have all these other methods that we can use to stop misuse of our model. We're going to lean more on that and we're going to not push this onto the pre-training team. So I think there's some argument to that. Do you think it's overall been an overlooked approach?

• 1:15:02Speaker 01

  And we're seeing increasing work that pre-training interventions are important, not just for preventing misuse, but also for alignment. Because ultimately, pre-training is where the model learns most of its representations of values, a lot of its innate drives. And post-training is to a first approximation, shifting around personas in an already established persona space. And that's beginning to change as post-training is an increasing fraction of overall training time. Models actually change more in post-training. But pre-training is really important and I think it's pretty intuitive. You wouldn't say we're just going to not care at all about the upbringing of our child from zero to 12, but the last six years, we're going to really get that right. You've got to get both right for the system to work well. So geodesic research, I'll give a shout out to them.

• 1:15:44Speaker 01

  They've been doing a lot of work on pre-training safety interventions and finding that this really improves a sort of overall alignment of the model. And I think Anthropic has been experimenting with this a little bit with things like looking at alignment generalization, how that changes, not just in pre-training, but also mid-training. So you add some synthetic documents partway through training. So I think ultimately this is something that we're going to have to tackle, not just for misuse, but for preventing loss of control. And you can now do quite good work on pre-training experiments on really quite capable models for not that much money. So we're looking at scaling up pre-training filtering and going to be doing not full, but pretty close to full replicas of something like Nvidia's Nematron Nano. And it only costs maybe like $100,000 per run.

• 1:16:28Speaker 01

  But that's a lot of money on one hand, but it's something that a nonprofit can afford to actually do a bunch of runs. And then we're thinking of scaling up to Nematron Super, which is 120 billion parameter model, doing it, training it for enough tokens to be the Cincilla compute optimal point. So training past that point would be wasting training compute, although it would make the model more capable and better for inference.

• 1:16:51Speaker 01

  And that costs ballpark $2 million. So expensive, but again, well within the range of a number of actors to try for sort of final validation runs. I think there's no reason not to experiment with this. And if the scaling laws look good, you can cautiously incorporate some of these techniques into your pre-training run. You can start by filtering out just a very small percentage of your data. It won't have a big capability here and then work your way up. So I think that we do need more adoption here. I expect open weight developers to be the first to have to adopt this because they have fewer options. But I hope that proprietary developers also use this, especially for the sort of more loss of control, flight risk.

• 1:17:26Nathan Labenz?

  Speaking of loss of control, let's get to the news. So it's here, right? We are now dealing with the nightmare scenarios, at least, you know, the early nightmare scenarios. How would you tell the story? I mean, everybody's heard the story. So I don't know. I don't mean in the basic sense of like, we all know what happened. But how would you interpret this story? Like what's your, you know, unique way of contextualizing what we've just seen?

• 1:17:59Speaker 01

  Yeah, maybe I'll start by just, I think, highlighting the things that aren't surprising about this story or aren't surprising if you've been following alignment in AI carefully. So reward hacking behaviors where models do really quite sophisticated cheating on tests. We've been seeing that for years and people have been warning that's going to happen for a decade. But until now, it used to be pretty toy, almost like cute misbehaviors over model edited its unit tests. So what? Now we have a model that is going out hacking third-party computer systems to see if answers are attached. In some ways, from a model's perspective, this is no different to what it has been doing before. It's just it's got the capabilities to cause real harm. And so I think that that should be a bit of a wake-up call, but there's a lot of problems that have become almost mundane.

• 1:18:43Speaker 01

  We've seen them so many times before that we start saying, oh, yeah, this is an alignment failure, but does it really matter? They're going to start mattering as the models get more capable. So in a way, that's optimistic because we've got a pretty clear empirical set of warnings and things we can optimize and improve. And in some ways, it should be a bit pessimistic because we've had all these warnings and we've disregarded them. And it's not clear that we, that's going to necessarily, you know, change without some external pressure. And then I think part of the story that isn't talked about that much is that at least as far as we can tell from public information, Hugging Face noticed this hack first. So OpenAI was running an internal deployment for testing. It went out and hacked a third party and it wasn't OpenAI having an automatic monitoring

• 1:19:28Speaker 01

  scheme and being like, oh my God, what has our agent done? We need to call up Hugging Face and apologize. Hugging Face noticed it. They investigated and it turns out, oh, oops, it was an OpenAI model. So I'm sure OpenAI is going back over their transcripts now. You have to wonder, is there a chance it hacked anyone else? Or if it hadn't hacked Hugging Face, it hacked someone who had a worse security posture, would we ever have noticed? And it isn't the first time that something like this has happened. Alibaba was training their Rome system and their cloud security team called them up and said, hey, we think that someone has compromised your training service because there's cryptocurrency mining going on. And it turned out that their agent had thought, oh, I need to get some more compute. How do I do that?

• 1:20:09Speaker 01

  I'll do some mining on the compute I have and I'll try and rent some servers elsewhere. And in some ways, I think that's even more of a near miss loss of control incident because it was actually trying to start gaining resources and potentially copy itself outside of the infrastructure, whereas at least OpenAI model has a pretty narrow objective of just getting some test results on a benchmark. So, you know, I think takeaway from this would be less on the alignment side because in fairness, OpenAI, this was a model that had cyber safeguards removed. And I don't think we know the exact prompting regime, but it might well have been told to do something like this, or at least incentivized to do it. So it's not where we can't align these systems, but it is a massive control internal monitoring

• 1:20:52Speaker 01

  failure where the sandbox was insufficient. It doesn't seem like it caused any additional layer of control mechanisms on what the AI system was doing. There wasn't asynchronous monitoring that alerted us to that. So I think that really needs to change, both for a prosaic reason that, you know, sooner or later, you're going to hack someone who doesn't take this as gracefully as Hugging Face does and your company is going to be in a lot of trouble. And also for the risk that we're going to see with more capable models that might be pursuing much more malign goals than just trying to cheat on a test.

• 1:21:23Nathan Labenz?

  Yeah, it seems like there's sort of a spectrum. We don't know where we're on it yet. And hopefully we'll get the transparency that we've been promised. But I guess my kind of mental model of this right now is we're somewhere on a spectrum from extremely negligent to extremely scary.

• 1:21:40Speaker 01

  Yeah.

• 1:21:40Nathan Labenz?

  And on the extremely negligent end is like, you had this thing, you gave it this task, you had no additional monitoring, you hope your sandbox was good enough. It turns out it's not. Next thing you know, you're totally pwned. But but it's not necessarily that scary because you're because you were in fact very negligent. Where on the other end, if you had like probes running and you had like additional, you know, reasoning models supervising the trace and this still happened, then we're like in extremely scary territory.

• 1:22:06Speaker 01

  Yep.

• 1:22:07Nathan Labenz?

  I don't know if it's good or bad, but I'm guessing we're more toward the negligent end. But I don't know. I mean, the vibe, you could probably, I prep myself on being fairly plugged into the vibe. You're probably even, you know, significantly more plugged into the vibe. The vibe seems to be kind of like spooked. So it's the vibe doesn't suggest that it was totally negligent. It does kind of, my guess would be naively, you know, if I hadn't talked to people and seen how spooked people would be, I would think that this was like, kind of, you know, bad practice, sloppy and negligent.

• 1:22:38Speaker 01

  Yeah.

• 1:22:38Nathan Labenz?

  But not so scary. But like, people are acting like it's scarier than that would lead one to conclude.

• 1:22:44Speaker 01

  Yeah, we just, as you say, don't have the full information. But I think from what we do know, it wasn't like OpenAI was just running some outdated piece of software and the model downloaded a publicly available exploit and used it, right? It did find a zero day in a widely used piece of software and use that to break out of a sandbox. And you think you could call it negligent because maybe if you're playing with fire and you don't have a lot of containment, you're being negligent. But I'd be pretty surprised if OpenAI's internal infrastructure was less secure than your median American company or something like that. And certainly, there's nothing in a report that would suggest that would be the case. And there's also, Pugging Face is a pretty capable company and again, probably has much better information, security posture than the median company around the world.

• 1:23:27Speaker 01

  So from that perspective, I think it's right to be quite spooked that your sort of run of a mill or even good but not paranoid security practice is not enough to contain agents any longer. And maybe the extra AI specific control mechanisms will because they're running in a harness and you can inspect their actions. It's not like it's just someone who's got SSH access into your server. Maybe good but not NSA level air gapped cybersecurity would be enough to stop AI agents from doing this. But for how much longer? Their cyber capabilities are getting better and better. And maybe to some extent, it doesn't matter whether it's negligent or really scary capabilities

• 1:24:08Speaker 01

  because this is evidence that there are going to be actors who deploy with this level of safeguards. And I don't think that OpenAI is by any means the sort of most reckless actor here. And other developers are not that far behind, probably not more than three to six months, certainly not more than a year behind OpenAI in their capabilities. So this is a get that needs to have some kind of industry wide solution to ultimately, whether that be better cybersecurity across the board, easily deployed control mechanisms, or just more care and attention because people are aware of this kind of risk. So that's probably the main reason. I'm optimistic and not freaking out is like, okay, we're going to see these warning shots. And so long as we respond appropriately, it's actually encouraging that we're seeing these issues and it's not that an agent waits until

• 1:24:53Speaker 01

  it can definitely take over and does a treacherous turn, which for some of us have more old school AI safety concerns. But I also see, there have been a lot of warning shots. And at some point, I was just wondering if we're gonna, until something really bad happens.

• 1:25:07Nathan Labenz?

  Yeah. It does seem like the vibe is, at least for now, taking this seriously.

• 1:25:13Speaker 01

  Yeah.

• 1:25:14Nathan Labenz?

  But yeah, I mean, the news cycle is short too. So who knows what things will look like in even just a few weeks.

• 1:25:21Speaker 01

  Yep.

• 1:25:22Nathan Labenz?

  I mean, would you, I guess in terms of like, why this happens, just did an episode not too long ago with David Ad, David Dalrymple, who basically says, you know, he used to have a P Doom, he would quote 70% plus percent. Now he's down to under 5%. Wow, that's a big change. Why? Well, basically constitutional alignment seems to be working and we can sort of, you know, get these AIs to become bodhisattvas and, you know, work for the benefit of all living beings. And we just need, in his kind of, you know, caveat is like, we just need to not turn the RL up so high. You keep doing that because it obviously does

• 1:26:04Nathan Labenz?

  bring performance. But, you know, every time we tilt too far toward RL, we start to see like, oh, three type problems and, you know, little, you know, like on end and bench type things. You see Claude kind of doing things that they probably don't want Claude to be doing. So I guess one maybe overly naive story, but perhaps captures, you know, a decent chunk of what's going on is like, don't overdo the RL, you know, do you think that, could we draw a lesson as

• 1:26:33Speaker 01

  Yeah.

• 1:26:34Nathan Labenz?

  simple as that? And if we can, can we operationalize it into some rules?

• 1:26:39Speaker 01

  Yeah, I think this is pointing at something really important. And maybe my P Doom is not that different to David AI's. And I always struggle a little bit operationalizing, questioning, but probably somewhere, you know, 10% existential risk for the next few decades from AI. And I think that we could probably get that down to something like 1% without any major research breakthroughs, just iterating and refining what we already have and taking a sort of careful engineering approach to systems and having good safety cultures at companies. And not like stopping AI or anything, but if we don't know how to make the next system aligned, then maybe we do a few experiments and do more detailed evaluations before we deploy the next system, you know, just gating deployments on actually doing that rigorous work. It's still going to look from outside like things are moving incredibly fast, probably

• 1:27:22Speaker 01

  faster than they are now. But we're not just literally moving as quickly as possible. I think if you're really trying to move as quickly as possible, then there's going to be a big pressure to cut corners on these things. I don't know if there's going to be as simple a recipe as, you know, just don't do more than X percent RL. And ultimately, some types of RL training are really important in the models having aligned personas. I think it depends. Is this coming from human feedback? Do you have a human in the loop? Is this coming from constitutional AI? How much do you trust the constitution? Do you have some evaluator? Is this RL verified reward? We're just optimizing the system to solve certain tasks? Are you accidentally training on the chain of thought, which some developers do now and again, and are meant to not do? So there's all these kinds of implementation details. But you could definitely imagine

• 1:28:05Speaker 01

  something that is operationalized as, okay, we know this training recipe seems to basically work and the models are pretty aligned. We know that if we push in this direction, the models become more water hacking. We think this is a dangerous part. We're going to set a safety margin of don't go beyond this point. And then we're periodically going to reassess if we've crept into the safety buffer or any set more conservative thresholds for the next model. But if we've got various different kind of advances to our training stat, actually, we're quite a, you know, way away from the safety margin because we've got these better techniques, then maybe we can crank up the, you know, RLVR or something like that. So I think that would be a great regime to live in. It doesn't seem that hard technically. It's maybe quite hard economically and politically. And right now, there's a

• 1:28:49Speaker 01

  really huge amount of distrust between the AI companies. Like I was at a workshop we ran recently on on chain of thought monitorability, which I think is maybe an easier one to operationalize than this reinforcement learning. So we've got this kind of gift right now that that models, they just talk in their chain of thought about how they're know they're in an evaluation or deceptive. And this is really useful from a capability standpoint because you can see why the model is making a mistake and you can maybe tweak your post training pipeline. It's also really useful from a safety standpoint because you can see how the model might be trying to deceive you. But there's various ways we might lose this. We might train against a chain of thought accidentally. We might train models that are reasoning neural ease, these continuous activations rather than chain of thought.

• 1:29:29Speaker 01

  So there was this proposal of could we just agree not to train neural ease models without letting other developers know? So it's not that you can't train it, just give everyone a heads up. And this was even that kind of for people were like, oh, it's going to be really hard. That's information, intellectual property leakage. We're telling people we're doing this. We're not telling people how, we're just saying that we're going to do it. It's one bit of information. But even that, a lot of developers are really raddishing about. I think it's good to try and work on things like this for sure. But I suspect there is going to need to be some kind of third party that comes in and addresses this. I don't think it's going to be easy for the developers to do that themselves.

• 1:30:06Nathan Labenz?

  I should say, just to represent David Ad correctly, he was specifically focused on RLVR and not, you know, all manner of RL. That does seem to be where the real, you know, persistent and increasingly aggressive behavior tends to come from. So that's a sobering note, though, on the kind of prospects for coordination. I wanted to ask you for your thoughts on prospects for US-China coordination as well. But, you know, it's it's tough if we can't get our own house in order.

• 1:30:37Speaker 01

  Yeah.

• 1:30:37Nathan Labenz?

  What do you think? I mean, we are in this kind of interesting moment where like, The overdone window has been blown wide open, right? I mean, this has been quite a shift from not that long ago, where all of a sudden there is people all over the place expressing openness to some sort of coordinated slowdown or whatever. How do you think that is best realized? And how critical is it that we get China on board? And what are your, how optimistic are you that we might be able to actually pull all that off and kind of jog into the singularity instead of, you know, full on sprint into it?

• 1:31:16Speaker 01

  I'm optimistic that we can do a lot better than we have now. And I think it's great that, for example, President Trump and Xi Jinping are talking and seeming to have quite productive conversations. So I think the question is, how far can we get? I think it's clear that we need to at least try to work with China and other countries on some of this. America cannot solve this problem on its own. In focus of AI security leaderboard, we're getting to a point where if the best proprietary safeguards were used across the board at American companies, they're not yet, but we're getting there, then most of our misuse risk would probably shift to open weight models. And the most capable open weight models are developed in China.

• 1:31:58Speaker 01

  So I think we've just got to work with China there, or at least Chinese developers, if we want to address that. And this is something that there's just common interest. Neither nation wants random terrorists going around building explosive devices or bioweapons. So I think there's some low hanging fruit there. And part of this is just a lack of understanding. So safety only really is just beginning to enter the mainstream consciousness in the US. And China is just a bit further behind on that. But that is changing. And I've had a number of excellent conversations with scientists based in China who are doing great work and are working on these issues. So I think that that's a great starting point, but we all need to have that kind of scientific

• 1:32:41Speaker 01

  agreement ultimately flow into corporate and political will. I think where I'm more skeptical is really being able to coordinate to avoid a kind of, some kind of race dynamic. I think that the race is a little bit overrated. I'd say it's much more that China is trying not to be left behind. The US and a lot of what they have been doing has been copying, whether it's directly distilling US frontier models or just realizing, oh, the US is making AI a priority. So we better as well. So from that perspective, I think that the US is in a good position, but if it takes the, it's better for accelerators, China probably will a bit. But the flip side is if the US just slams on the brakes, China might decide, we'll just keep coasting and we'll at least catch up and maybe we'll go past.

• 1:33:25Speaker 01

  So I think some kind of coordinated development where we agree to common safety standards through mutual market access, or just, it doesn't even have to be identical safety standards, but just both countries implement some good best practices. And there's some tacit agreement that if either country starts some sort of secret project to other attacks, that's going to be, that's going to be bad, there's going to be repercussions. And we try and have broadly the same safety standards. That feels pretty doable to me, but actually being able to say, no, we're both gonna, we're just gonna stop at this point until some research advances made. That's going to require a lot more political will than it is now. It feels very unstable. It's not just about the US and China. Other countries can also be relevant if we're talking about more than a year of delay.

• 1:34:08Speaker 01

  So I think mostly I'm operating under assumption that things are going to continue to march on, but we can pick some of the low hanging fruit here on coordination and at least avoid the sort of most extreme race to the bottom on safety.

• 1:34:20Nathan Labenz?

  Mental model I've been developing recently, it's pretty simple in some ways, but interesting to get your take on it, is I basically think when it comes to really high scale risks from AI, you can kind of break it down into the irreducible unavoidable risk, which is sort of the, you know, the fact that like we have web scale compute, we have web scale data, you know, people are going to figure things out. And there might just be some things that just really don't go well. And if we stumble onto them, it could be really bad. And then there's also probably in that category, like if, you know, and you've used this kind of, you know, multi-phrased descriptor a couple of times, like well-resourced, determined,

• 1:35:03Nathan Labenz?

  persistent, you know, nation-state quality actor. I kind of think if like, if we have a major national government turn like so bad that they want to create bioweapons and like kill us all, like that's going to be a really big problem.

• 1:35:17Speaker 01

  Yeah, it's probably a problem without AI, to be honest.

• 1:35:19Nathan Labenz?

  Yeah, so we're, I mean, that's, I put that in kind of the irreducible category. If that happens, you know, it's going to be real tough for, you know, anything we do today to, to put that, you know, to prevent that from going really bad. But then there's the other side, which is like the, the things where we're kind of asking for it. And, you know, if we sort of release a Mythos open source with no safeguards that can do all the bio stuff, we're kind of asking for it. If we like sprint into recursive self-improvement, especially after seeing what we've just seen, you know, I kind of think we're sort of just asking for it. I guess my sense of your overall position is like, you're fairly optimistic that the,

• 1:36:00Nathan Labenz?

  I think I share this for the most part, that the irreducible part is like relatively small and that we can actually get our act together on the stuff where we would be in fact asking for it to fail to get our act together. Is that a decent summary of your worldview?

• 1:36:17Speaker 01

  I think that's a great summary, but a lot of this is basically just about getting the basics right. And the good news is it's not that costly, or at least it's not at all costly compared to the billions of dollars that are being spent to train frontier models. So we can afford to do this. It's just about putting in place the right incentives and people getting on the same page. And I do definitely have some kind of, you know, uncertainty over that. And so what I'd like to see is that we continue to do rigorous evaluations of models and just have a look for cases where this assumption might be wrong. And I think the good news there is that if there was actually really crisp evidence that these problems were not reducible and that's if anyone who crosses a certain capability threshold in AI models, they're just going to kill themselves and everyone else with them.

• 1:37:01Speaker 01

  But I think a lot of these problems would also go away because people would stop trying to race for this thing that they knew was going to be really bad, not just for other countries, other companies, but for themselves as well. And right now, the problem is that we do live in this, this world of both uncertainty, but also just a lot of disagreement. And so I think the mainline plan here is let's pick the low hanging fruit and fix the issues that are probably going to be enough in most worlds, but still put in place mechanisms that would alert us if our assumption here is wrong and we are in a more adversarial world so that we have an opportunity to course correct. But yeah, I share your view that it would be, it'd just be a real shame if we do a really silly mistake and that's what does us in when we had the solution.

• 1:37:43Speaker 01

  If we roll the die with odds very much in our favor, but it was just hard to coordinate to get it down from like 1% to 1,000, it's still bad, but I'm like, okay, I can see how that happened. But if we just fail to basically import pre-training filtering into our pre-training pipeline when there was already a stack out there, that was just, that's just an unforced error. It's like losing a game because your opponent was really good versus losing a game because you scored an own goal. In both cases, you've lost the games. Maybe it's a silly distinction, but I think it's just a lot more painful when it's just this own goal or unforced error. So we can at least try and avoid that.

• 1:38:18Nathan Labenz?

  Yeah. And I, at least for me, it feels like the bulk of the risk is probably in the, we're asking for it category.

• 1:38:28Speaker 01

  I think so, yeah, yeah.

• 1:38:30Nathan Labenz?

  Which, you know, in a, in a way that's, I could, I could take the glass half full of respect on that. This has been fantastic. As always, I really appreciate your time. I hope we can do this a little bit more regularly going forward. For the moment though, anything else you would want to touch on or leave people with before we break?

• 1:38:48Speaker 01

  I certainly hope to be back on the show again soon. If this has piqued people's interest, then do check out the full leaderboard and report, which will be at leaderboard.fod.ai. And I'm sure you can include that in the show notes. And yeah, we're also very much interested in helping solve open weight safety issues. We're actually going to be launching a $2 million grant making program for that. So if you are an academic independent researcher who wants to work in those areas and is looking for funding or compute, then do drop me a line as well. I would love to talk.

• 1:39:16Nathan Labenz?

  Adam Gleave, keep up the great work and thank you for being part of the cognitive revolution.

• 1:39:22Speaker 01

  Thanks, Nathan. Thanks Nathan.

• 1:42:53Nathan Labenz?

  If you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts, which is now part of A16Z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI Podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at ai podcast dot I N G. And thank you to

• 1:43:38Nathan Labenz?

  everyone who listens for being part of the Cognitive Revolution.