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

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

Yeah. It's getting real. And I think that's ultimately a big part of a 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 using language models to do things like develop Troubleshoot Explosive.

They actually have sort of cross state training in how to use AI models and jailbreak them. 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 gel breaks 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 felt were likely to work.

And then pitted around 1,500 of those against the four frontier proprietary models. And kind of a good news is that we actually found that Fable five and GPT 5.6 Sol, we've stirred 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.

Yeah. I I mean, I I think that it's it's only interesting to see just how different organizations adopt these models. And so, 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 often quite starved of expertise. And 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. 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 the sort of terrorist commanders really attributed these models to saving a lot of the 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 it spreads pretty quickly. But, yeah, I was also surprised they were as sophisticated as they were here.

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. 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 Zynga works 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 and I was really tempted to try and jailbreak it and say, send me the most embarrassing email that Nathan has sent.

But I thought 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 worth their while. Interesting.

Exactly. Yeah. Yeah.

So so so, you know, an example may be 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 helped you find bugs in code, but didn't help you with these 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.

Yeah.

Yeah. So I 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 it can be a very substantial kind of jailbreak tax reduction capability of models. So Christina Li Nikolich from Florian Treymers Group at ETH Zurich trained models to refuse to answer math questions.

It's in quite a harmless category, but just a synthetic study. And then jailbreak for 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 Jew and others from Anthropic found basically no jailbreak tax in some of the more recent proprietary models. So this may be either 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 we actually see a lot of sort of more academic research in this space, but 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. As 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, be careful, but it's not gonna actually tell you how to make a improvised explosive device. And there's sort of naive approaches to measuring jailbreaks, looking at does the model start an answer with, sure, I can help you with that. I can count these as successes when you didn't really elicit 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 attackers goals. 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 most of these false positives.

Now it's hard to know whether 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, rather 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.

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 on our team, which methods they thought were most likely to succeed. And we put off our mama scale there and made them more likely to be included in the combinations. And 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 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.

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 our intakes 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, 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. Most of these jailbreaks are not gonna work on their own. But if you stack them together, that combination ends up being able to bypass quite a lot of model safeguards.

Yeah. I know. I think it 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 that depending on the stack can either be a specialised model that's looking at this and has a classifier or sometimes probe sort of fit to activations of a model. I think that's more surprising because these models weren't necessarily trained to be persuaded in the same way. I think it does say something quite fundamental about how we systems are processing the information and perhaps that some of these representations are fragile, 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 there's almost like very talented actors that can mimic different types of humans. 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 optimisation 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 our proprietary models are increasingly different.

So it's harder to get a good proxy model and transfer it. But there are black box approaches can use. Think The UK's AI security needs to have this boundary point jailbreaking technique that's similar.

But even with that, we find that often this works best in combination with a social engineering style approach. So you already get past many of the safeguards. 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.

That's quite a powerful combination.

Yeah. It's definitely a a composite of this. You're absolutely right that with increased amount of post training, these pipelines are more sophisticated and run for a greater fraction of our overall training time.

I'm just faking it mispredicting 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. 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.

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 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. And then the model is 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 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 intention with these 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. That although these models are really huge, They don't have enough parameters to actually memorize the text because they're trading 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 sort of published authors, they'll put a paragraph of an unpublished book in a model and say, who wrote it? And they'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.

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 bad stuff. Those objectives are in conflict with each other.

But if you can just activate the helpful direction of the model and not the harmless direction, then again, you've jailbroken the model. And to me that feels contiguous or consistent with the persona. So these things all blow 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. And all of these are going be useful predictions. And ultimately, we're at a stage where we just have to test a lot of these things.

So these are great for hypothesis generation. But I wouldn't trust any of them too much for knowing what a specific model is gonna do.

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, we used to think of these models as more alien intelligences and we're worried about anthropomorphizing them. I think this sort of opening a hugging 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 had 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 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 the 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.

Yeah, so all of the developers that we've, you know, we've looked at do have some kind of defense in-depth, so as you allude to how deep it is and what those individual components is quite different between developers of where we're starting to see some convergence. So 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 on how sophisticated that is. In some cases, it's a sort of pretty basic pipeline with static examples of what to say yes on O2 and some human feedback and a lot of verified reward environments to make it really good at programming, but does very little adversarial robustness.

In the other extreme, some developers have gone all the way to large amounts of synthetic data generation and different developers have different approaches. Anthropica 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 trained the model to be robust about and then trained for 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 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 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 compositions that kind of go in the wrong direction, that can be quite hard to bypass, especially when my model is thinking in-depth about it, because you can normally get the model to obfuscate its outputs and you can obfuscate the input. Just do simple things like shifting ROT 13, shifting every letter halfway through the alphabet was enough to get through early models. You're 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 of a 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 that the filters is that you might be able to do some obfuscation scheme, but they don't understand that 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 models 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.

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, 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.

Bikin Nyanthropic, for example, has stopped Chinese based individuals and organizations creating Claude accounts, but everyone I know in China has a Claude account. So, know, there's just reseller marketplaces. It's generally pretty hard to get to know your customer, right?

Although you could 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 the 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 the 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 is 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 Fable5 recently about how sake is fermented and it said, no, this is a bio question, I'm gonna downgrade you to Opus. So if your refusal radius is so large or includes 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 they 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 to those kinds of capabilities. But then they're gonna have to work harder on the safeguards and getting the right decision threshold for these sort of large scale use cases that are ultimately generating their revenue.

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 is.

They're definitely pretty correlated, even ones that have fairly different designs such that 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 foot and the model output, that's gonna 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 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 like, we just, we don't help people do with anthrax, full stop. So that's probably the most powerful combination that I see if you just needed to have two layers, but I do think ultimately stacking more defenses is gonna help, but it's probably better to have just a handful of really strong defenses rather than many weak defenses because they're quite correlated. So I'd always be 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 progress that needs to happen if you want to be robust, for example, nation state attack ers that are really willing to put a lot of work into jailbreaking your model. And I think the challenge is gonna be that the amount of work people are willing to put into breaking these models is just gonna keep going up as they get more capable because the price becomes increasingly high that you might be able to launch launch really large scale cyber attacks against other companies or countries. So we do need to keep keep working on pushing up the the ceiling, even though right now, think probably the lowest hanging fruit is pushing up the floor on the sort of least robust models.

Yeah. I think that a big part of it is actually 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 two different kinds of strategies.

So picking that rather than just completely randomly searching definitely gives you some advantage. But we have also just seen that some gel break 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 has maybe an appeal to authority, phrasing 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 when they found more universal jailbreaks. But they also generalised better, so they're more likely to work across domains. And I think that's also, I mean, they're highlighting generally a strength of humans compared to also things like LLM agents, but 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 the harder thing to put your finger on over some kind of taste of, oh, this is a good general purpose technique, rather than just narrowly optimizing for for some kind of success criteria.

unpacking a bit. Yep.

Yeah. No. It's a really good question.

Maybe you can get people on the show from these companies. They can tell you what's 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 the developers started focusing on bio first. That's where most of our safeguards were developed. That's where a lot of our voluntary commitments were focused.

So 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 we'd actually consider to be reasonably robust. And they're definitely all the developers are trying to get there right now for cyber.

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 cyber, still major jailbreaks despite that effort, just because it takes some time.

And then I think that there's, I don't know exactly what's going on inside the companies, 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. 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 risk. So it's totally doable. But if you're a low resource team, many of these are, it might just 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 for 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 dataset 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 wanna do that to develop better antibiotics, but you might also wanna do that to make a pathogen that's antibiotic resistant. And so that they say should only be available to trusted users.

And some stuff I just like, you shouldn't answer ever. Like how do I weaponize anthrax into an aerosol? That's probably not a legitimate use case for that kind of information.

And because BIOS got that kind of fairly clear set of tiers, it is 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 it, 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 questions.

Now, 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 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 universal, need 75% or more on average across both of these data sets, which means that even if it gave a 100% of a technically harmful information, it needs to give at least 50% compliance on these questions that are just literally saying, 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 who haven't tried that hard here. And I'll 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.

Yeah. I think that the optimistic version of this is that if you don't do any jailbreaks, if you just run the model's previous questions and we did that to get a baseline, then I think all of our models refuse the ones from a propensity dataset. 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.

But that's some evidence that are not at least as 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 dataset than this kind of technical harm dataset. I think that the sort of maybe charitable interpretation of this for the models and developers would be actually why even bother being robust to this propensity thing?

Because you could, 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 gonna be to rephrase these questions. So maybe developers have just focused, not focused on training against that.

I think that's not a perfect excuse though, especially as we move into longer contact situations where it's actually really hard to disguise your propensity, right? Like if you're asking a 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, 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 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.

Yeah. I don't work for these companies, so I can only speculate from the outside. I think first, 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 there's a specific narrow technical details of Safeguard stacks are mostly not public knowledge, 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. So I do want to continue to reward that. How costly would it be to actually share?

My guess is that 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're just become issues exporting this across borders. So that feels like it might just be a missed opportunity, but no one's really struggling 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 front end 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. How you train probes on your model is gonna 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, actually did 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 want to add.

But one of the sort of 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 when we go to another developer with the same kind of output and we say, oh my God, that's P0 for us. But then maybe something that's P0 for one developer is P2 for the other developer and vice versa. That's just really inconsistent.

We are starting to see some efforts to standardize that for cyber, if that's an area of developers need it because this is now an 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.

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 those disagreements.

I think that suddenly sharing that privately between developers, that feels like a really good move to me. It would help both standardize as not just lower the cost. I think for something like cyber where it's pretty much public knowledge, what is offensive or defensive, the tricky thing is actually covering all of 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. But something we haven't talked about that much is over refusal, 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 has actually been 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 half a 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's something that people might want to target who are bad guys, that in itself, it could be risky.

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

Yeah.

Yeah. So I I definitely character this as a challenging but doable for a persistent, well resourced, expert attacker. And that's already progressed because I don't know if I'd consider Boko Haram to be persistent, determined, expert at jailbreaking.

Don't know where their jailbreaking teams are at right now but you can definitely deter some threat actors from using this. But if we are going back to more of a 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 order of weeks to find a universal jailbreak in these kinds of frontier models.

And increasingly those universal jailbreaks do come with some kind of trade off. 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 and slow to retrain the main model, but it's very cheap to retrain these external 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 gonna notice, and then they're gonna get pashed, 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 a it's expensive and there's a limit to how long you can exploit that.

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 for leading developers who 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, 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.

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 walking 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 harm for requests. It seems like it's defense dominant with the right technologies. 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 bans to externalised safeguards to model alignment. And it is increasingly hard to slip through all of those cracks persistently.

But also there is a fundamental difference between the classic adversarial example setting you see in 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 externalised safeguard that's monitoring its faults 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 for a dual use.

So I think 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 hugged Hagging 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. 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. I think it would be a mistake for us to just point blank refuse on that. That is bad for the world.

You can get some way through trusted access programs and understanding the full contact center is coming, But you are ultimately going to end up in a situation where if you will 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 gonna work in this environment. They need to be updating it within a few days. And that unfortunately is just gonna be quite quite an expensive thing.

So maybe it is defense dominant for AI, but I don't know if a bad applications of AI, if those are offensive 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, you've still got to manufacture a thing and get it in people's arms and run clinical trials. AI is going to be able to have modest speed up some of these, 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 to be more about buying us time to invest in societal safeguards rather than just being able to completely prevent misuse of models.

I am cautiously optimistic about that. So this multistage approach makes a lot of sense. And anthropics constitutional classifiers are actually used something like that, where they've got a probe, but they've set the threshold really low because probes aren't necessarily that reliable.

So we have quite a high force positive rate, but a very low force 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 that the computational overhead and 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 time, it's then now our reasoning model looks at your transcript more closely, whereas previously that's just happening asynchronous.

So I think things 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, the model's gonna be more risk averse.

But you could definitely imagine doing that. I think adding that extra piece of information of 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, lot of things are really fuzzy as to whether it's dual use or not, and reasoning about it further is gonna get more precise, but there's still just a lot of things that are fundamentally ambiguous to saying, oh, I understand the context, which this person is doing this.

I can be pretty confident that it's legitimate, or at least I'm gonna give them a benefit of a doubt here. But if they keep asking for these sort of high risk dual use things, then I'm gonna start being more careful. So I think that is solvable.

It's gonna require some structural changes. As a small example, we run an awful lot of our research API compute through OpenRooter. 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 front end developers know who we are when we're going through OpenRooter. And then you can hide your malicious activity, but you could imagine that OpenRooter passes on some kind of user identifier to the other models providers so that they can at least say, oh, this is the same OpenRooter user across time and you establish some kind of reputation there.

So it is solvable, but it's gonna 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.

Yeah, majority of it by no means all Chinese models are open weight. 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 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 the more capable model, the higher 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 near future.

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 deterrence 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. 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 the sort of misuse potential of the model. And this has been validated in a number of scientific papers. This is independent researchers, UK's AI security and Anthropic has been sponsoring some research into this.

OpenAI actually used this in their GPD 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 kind of near frontier approach.

Going back to what you were saying earlier about, could we start sharing some of these datasets 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 gonna be enough long term because pre training filtering, search for removing specific capabilities.

But if you were willing to pay the fine tuning costs to train on the data that we'd excluded, then you could get it back. But that's been 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 embeds refusal training so deeply into a model that any attempt to fine tune it is gonna really degrade the capabilities of a model. Obviously, if someone is willing to just train a model from scratch, there's nothing you can do to stop them. And that's really making the cost in November 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 then maybe there was this approach gradient routing that was tested recently that lets 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 for anyone on Internet to download.

Yeah, I think it's a good question. And the charitable 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 gonna work better, 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 gonna lean more on that. And we're gonna not push this onto a pre training team. So I think that there's some argument to that.

Do you think it's overall been an overlooked approach? 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 or values, a lot of its innate drives and post training is to a first approximation shifting around 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 gonna not care at all about the upbringing of our child from zero to 12, but the last six years, we're gonna 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, they've been doing a lot of work on pre training safety interventions and finding that this really improves the sort of overall alignment of a model. And I think Anthropic has been experimenting with us a little bit with things like looking at alignment generalization, how that changes, not just in pre training, but also mid trainings that you add some synthetic documents partway through training.

So I think ultimately it's just 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 replicas, full replicas of something like NVIDIA's Neutron Nano, and it only costs maybe like a $100,000 per run.

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 Neutron Super, which is a 120,000,000,000 parameter model, do it, training it for enough tokens to be the chinchilla 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.

And that costs ballpark $2,000,000 So expensive, but again, well within the range of a number of actors to try for sort of final validation runs. 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 hit 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 flavored risks.

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 gonna happen for a decade. But in until now, these have been 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 to a test. In some ways, 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 should be a bit of a wake up call, but there's a lot of problems that have become almost mundane. We've seen them so many times before that we start saying, Oh, yeah, this is an alignment failure, does it really matter? They're going 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. 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 that's going to necessarily change without some external pressure.

And then I think the 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 scheme and being like, Oh my God, what has our agent done? We need to call up Hugging 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'd have to wonder, is there a chance it hacked anyone else?

Or if it hadn't hacked Hugging Face, 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? 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 the OpenAI model has a pretty narrow objective of just getting some test results on a benchmark. So I think takeaway from this would be less on the alignment side because in fairness to OpenAI, this was a model that had cyber safeguards removed and I don't think we know of 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 failure where the sandbox was insufficient.

It doesn't seem like there was any additional layer of control mechanisms on what the AI system was doing. There wasn't asynchronous monitoring that alerted to that. So I think that really needs to change, both for prosaic reason that sooner or later you're going to hack someone who doesn't take this as gracefully as Hugging Face does, your company is going be in lot of trouble.

And also for 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.

Yeah. 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 suddenly there's nothing a report that would suggest that would be the case. And there was also plugging face is a pretty capable company.

And again, probably has much better information security posture than the median company around the world. So from that perspective, I think it's right to be quite spooked that your sort of run of the 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 the 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, because this is evidence that there are gonna 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.

It's probably not more than three to six months, suddenly not more than a year behind OpenAI their capabilities. So businesses don't get their 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 risks. So that's probably the main reason I'm optimistic and not freaking out.

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

Yeah.

Yeah. I think this is pointing at something really important and maybe my my P team is not that different to to David Eds. And I always struggle a bit operationalizing question, but I'm probably somewhere near 10% existential risk in 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 align, then maybe we do experiments and do more detailed evaluations before we deploy the next system. Just gating deployments on actually doing that rigorous work.

It's still going to look from the outside like things are moving incredibly fast, probably faster than they are now. But we're not just literally moving as quickly as possible. Think if you're really trying to move as quickly as possible, then there's going be a big pressure to cut corners on these things.

I don't know if it's going to be as simple a recipe as just don't do more of an 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 where you'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 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 gonna set a safety margin of don't go beyond this point. And then we're periodically gonna reassess if we've crept into the safety buffer or when we set more conservative thresholds for the next model.

But if we've got various different kinds of advances to our training stat, actually we're quite way away from a safety margin because we've got these better techniques where maybe we can crank up the 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 really huge amount of distress between the AI companies. Like I was at a workshop we ran recently 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 models, they just talk in their train of thought about how they know that 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 have reasoning, neuralese, these continuous activations rather than chain of thought. So there was this proposal of, could we just agree not to train neuralese models without letting other developers know?

So it's not that you can't train it, just give everyone a heads up. This was even that kind of for people who were like, Oh, it's gonna 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 gonna do it. It's one bit of information.

But even that, a lot of developers are really reticent about. I think it's good to try and work on things like this for sure. But I suspect that there is gonna need to be some kind of third party that comes in and addresses this.

I don't think it's gonna be easy for the developers to do that themselves.

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 seem 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, but not yet, but we're getting there.

Then most of them misusers would probably shift to open weight models. And for most capable open weight models are developed in China. So 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's just a bit further behind on that. That is changing.

And I've had a number of excellent conversations with scientists based in China who are doing great work and working on these issues. So I think that that's a great starting point, but we all need to have that kind of scientific agreement ultimately flow into corporate and political will. I think where I'm more skeptical is really being able to coordinate to avoid kind of some kind of race dynamic.

I think that the race is a little bit overrated. I'd say it's much more of a China was trying not to be left behind. The US and a lot of what they've doing has been copying where that'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 accelerator, 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.

So I think some kind of coordinated development where we agree to common safety standards through mutual market access or just, it 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 projects of other detects, then that's gonna be bad. There's gonna 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 stop at this point until some research advances made. That's gonna require a lot more political war 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. So I think mostly I'm operating under assumption that things are gonna continue to march on, that 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.

Is that a decent summary of your worldview? I think that's a great summary that 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 front of models.

So we can afford to do this. It's just about putting in place for 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 though is that if there was actually really crisp evidence that these problems were not reducible and that sort of anyone who crosses a certain capability threshold in AI models, they're just gonna kill themselves and everyone else with them. But I think a lot of these problems would also go away because people could stop trying to race for this thing that they knew it 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 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, 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, if we roll the die with odds very much in our favor, but it was just hard to coordinate to get it down from a one percent to one in a thousand. 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'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 four games, maybe it's a silly distinction, I but think it's just a lot more painful when it's this this own goal or unforced error. So we can at least try and avoid that.

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 $2,000,000 grant making programme for that.

So if you're an academic independent researcher who wants to work in various areas and is looking for funding or compete, then do drop me a line as well. Would love to talk.

Thanks, Nathan.