00:02:47 · Max Justicz
[00:02:47] S0: I've been wondering that as well. I have some ideas. For one, it does seem to me like the frontier models are pretty effective at locking down access to some of the most, like, egregious vulnerability hunting requests. I was using Daybreak Blue, which is kind of the easiest to access elevated security permissions model that OpenAI offers. And even since finding this bug, they've really locked it down. Like, the same requests that I was making to find this bug are no longer working. So it does seem like the labs are at least getting a little bit better at locking down these models. I I haven't really been able to get Claude to do anything. It's super useful for me. That said, you know, I I also spun up an obliterated GLM 5.3 model just to see how it fared. And it it it was pretty good. Like, it it it was definitely a bit dumber than the g b d 5.6 models that I I was also playing with, but certainly powerful enough to to be dangerous and find kind of the the more shallow bugs. As for why we haven't seen more harm with this, I think... I wonder if we're rate limited on some of these harms just by the the number of actors actually trying to to to do harmful things. Like, the folks running ransomware campaigns are probably still running ransomware campaigns, and and maybe they're able to find some new targets. But, like, really, what it comes down to is I think that it just hasn't really been long enough yet. I think the tools just aren't in enough people's hands to to have caused major incidents. But I think we'll see them. I think it's kind of just a matter of time. Right.
00:04:37 · Ruby Justice Thelot, Max Justicz
[00:04:37] S1: So how long does it usually take to go from finding this bug to fixing it or bugs bugs like this? [00:04:49] S0: I think people are definitely just the rate limiting step there. I had reported this bug actually back in September. So it took five or six weeks for them to fix it and roll out a patch. But I think during that time, the maintainers were literally on vacation for part of it. So I think a patch could have been written and released later that day. So I I think the patches can come out quickly. It's really just a matter of having people, like, deploy them as urgently as they need to. [00:05:18] S1: So what should people who maintain packages be doing right now in light of all of these cyber incidents and cyber risk for models? [00:05:27] S0: It's very easy to sign up for Daybreak Blue or Anthropic. I think they call it the cyber... What is it? Cyber verification program or something like that. And you should just sign up and and aim it at your code base and and help it... Help you audit your your own code. I think that's kind of just table stakes now. And if you really wanna get fancy with it, I would spin up an obliterated GLM 5.3. And since there's a a startup called, I think, Obliteration dot ai [00:05:57] S1: Yeah. We needed the CEO. [00:05:58] S0: Oh, really? Yeah. I messaged him on LinkedIn the other day just to say thank you for making my life easier. But, yeah, I would recommend throwing $100 their way and aiming it at your code base as well. It's probably gonna be more effective.
00:05:58 · Max Justicz, Ruby Justice Thelot
[00:05:58] S0: Oh, really? Yeah. I messaged him on LinkedIn the other day just to say thank you for making my life easier. But, yeah, I would recommend throwing $100 their way and aiming it at your code base as well. It's probably gonna be more effective. [00:06:11] S1: Do you think that companies like Obliteration and those that offer obliterated versions of very cyber capable models are net good or net bad for the world? [00:06:20] S0: I think in the longer term, net bad to be providing kind of misaligned AI as a service. I think in the short term, it's actually good in that it's going to allow defenders to audit their code more quickly. So I think I think right now, it's... There's also just weird, like, second order effects. Right? Like, I was able to use that service in a way that might be raising some awareness to these issues. And I think that has a significantly good impact right now. So yeah, I think in the long term, the normalizing, providing access to AI models that are willing to do harmful things is is probably not good. [00:07:14] S1: Right. That makes sense.
00:07:14 · Ruby Justice Thelot, S2
[00:07:14] S1: Right. That makes sense. Ad[00:07:17] S2: I wanna shout out our sponsor, Lovable. You know, the app you've been meaning to build? The internal tool, side project, or product you'd otherwise lose a weekend to? With Lovable, you get software you can ship today. That includes an editable code base you can inspect and change, plus two way GitHub sync. Lovable also handles the backend infrastructure, including managed Postgres, auth, storage, hosting, payments, and more so you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable. Dev. Now on to the episode. [00:07:56] S1: So is there any way in which this should affect the things that ordinary users do with their computers to stay safe on the Internet?
00:07:56 · Ruby Justice Thelot, Max Justicz
[00:07:56] S1: So is there any way in which this should affect the things that ordinary users do with their computers to stay safe on the Internet? [00:08:06] S0: Unfortunately, I think ordinary users are in a pretty bad spot right now. I don't think there's much you can do as a user besides, like, use fewer services. I mean, I I would stick with my advice pre AI security apocalypse, which is, you know, stick with major smartphone brands that are likely to get patched more quickly, you know, use major operating systems, keep things up to date. But it's a little bit futile right now. One of the other projects that I took on recently, actually, after finding this bug, was I just bought a modern fully patched Android phone, aimed the same models at it, and very quickly got kind of the worst of the worst finding where I could send a malicious message, just over cell networks to an Android phone and and take it over, run arbitrary code. And the patch for that one's still being rolled out. But, you know, we're in a world where me, as someone with a full time job not doing security stuff, could just kinda, like, dick around one afternoon and find that in a few hours and exploit it. And, yeah, there's there's not really much you can do as a normal person to protect yourself against that right now. Yeah. But, like, how [00:09:18] S1: good are you? Like, could anybody point models at software like you did and get similar results? Or is there some sort of cybersecurity expertise that you have that makes it easier for you to do that?
00:09:18 · Ruby Justice Thelot, Max Justicz
[00:09:18] S1: good are you? Like, could anybody point models at software like you did and get similar results? Or is there some sort of cybersecurity expertise that you have that makes it easier for you to do that? [00:09:29] S0: Well, this is why I think I'm kind of an interesting case study here. I think that I'm okay. I'm not one of the world's best hackers or anything, but I'm someone who had some success doing bug bounties. And I've seen just a massive uplift in what I can do because of these models. And I think that, like, marginal difference due to my skills is just gonna quickly diminish to the point where, you you know, if if you're completely unskilled, you'll be able to do the same thing. Like, I have a little bit of taste in, like, being able to aim the models at at things that I think are worth looking at. But, like, I mean, what kind of mode is that? Like, I I I don't think that's gonna last very long. So I think I'm kind of on the bleeding edge of... Like, think I view this work as kind of like a canary. Like, I think that, you know, I have a little... Like, you know, probably a couple months ago, like the most skilled hackers were seeing additional results due to this. Now I am. And pretty soon, just normal computer users will be able to find stuff like this. I'm pretty confident. [00:10:38] S1: Mean, will they or will the software maintainers have already found it and patched it first?