XBOW's AI found a full-chain Chrome vulnerability affecting 1 billion users in just 6 days

Speaker 1: That's right. And CEO. Welcome to the show. How are you doing? Here he is.

Speaker 2: Thank you so much for taking the time. Hey.

Speaker 1: What's happening?

Speaker 7: Welcome. Thanks very much for for having me.

Speaker 2: Thank you for joining. Yes. I would love to to get a brief introduction on yourself and the company first, and then I wanna go into just cybersecurity and what you're seeing, what is interesting, how big is the problem, how effective is the solution.

Speaker 7: XBO is a company that automates security research. So it does the work of an offensive security researcher. You point it at a target and it will report the vulnerabilities it finds working like a human expert. Yeah. Now, today, we are announcing a particular vulnerability found by OG. A full chain vulnerability in Chrome.

Speaker 2: Okay.

Speaker 7: What that means is that you just point your web browser at one malicious website, and the attacker will own your entire Windows PC. Okay. Your email, bank account, passwords, key... Keystrokes, and so this vulnerability affects a billion users. Wow. Google Google acknowledged the problem, paid their maximum bounty. But what's really interesting is that it only took six days.

Speaker 2: Okay.

Speaker 7: Six days by a very skilled security researcher. Yeah. It was joined by a swarm of agents. Sure.

Speaker 2: That makes sense.

Speaker 1: Yeah. And the other the other thing that's notable is, like, you know, finding this this exploit that that would impact a billion users in a very meaningful way. And and congratulations. Your prize money is $250,000, which

Speaker 2: Yeah. Do prize monies need to go way up so you can offset compute? I don't know what you spend. I don't know if you want to share, but I've heard of, like, million dollar agent swarm runs, essentially prompts that wind up working for days of wall clock time, and they're burning GPUs for hours and hours. And it feels like it can get expensive. But the value of catching these bugs is so high, especially right now that should companies be spending more on these, like, white hat bounties?

Speaker 7: So definitely, everyone should get their defenses up. But, no, it was not that expensive.

Speaker 4: Okay. We

Speaker 7: used an off the shelf model, a GBT five six Soul.

Speaker 2: Wow.

Speaker 7: And it only cost $40,000.

Speaker 2: Okay. Wow. Yeah. And and so how how how integrated are you with OpenAI to be able to do offensive cybersecurity testing? Do you need to get whitelisted with them in some way to to actually be able to run prompts that might get rejected just by the default ChatGPT front end?

Speaker 7: So OpenAI are great partners of ours, as are all of the other front end apps. Yeah. In this case, it was a publicly released model, but we are part of daybreak, and

Speaker 2: Yeah.

Speaker 7: Of course, we also got to got to play with the restricted models. Yeah. However, we really wanted to make the point here that it's possible to do with a publicly available model.

Speaker 2: Yeah. Yeah. That makes a lot of sense.

Speaker 1: So, yeah, break break this down a little bit more because because... So so the browser, like, you would simply have to navigate and land on the website. You don't need to you wouldn't need to to download anything or run any software. It's literally just you land on it and it's over for you. Is that is that generally correct?

Speaker 7: That's exactly right. So this this type of bug in Chrome is really really hard to find. Every major security intelligence agency in the world has a team of people doing this kind of thing, and some of my contacts have been telling me that it takes months, something like six months to do this with a team of people. Yeah. And this time, it was done in just six days by one person. I mean, not really building on our previous experience that he had gathered with his with his with his colleague.

Speaker 1: Yeah. When you find an exploit like this, does it ever go through your head like maybe somebody would have found this before but wouldn't have the incentive to share it publicly or obviously they Yeah. Like a malicious actor could make a lot more than $250,000, like, exploiting this, then... And and so it begs the question, yeah, was this already being abused for a while?

Speaker 7: So we have no evidence of that. I don't know for sure, but don't know whether someone else has found this particular bug before. But what the story does tell you is that everybody really has to get their defenses in order. Cybersecurity is now right at the top of what everyone needs to be worried about. Take action. Find the bugs before the bad guys do.

Speaker 2: Mhmm.

Speaker 1: What what are your predictions for for the next three to six months? Like, feels like the... This... Everyone inside of tech is is is well aware of the issue and and, you know, making sure that that their walls are are are as strong as possible. And it feels like the the the sort of mythos moment really helped bring this sort of risk into even outside of the tech industry. Let's call it like the the Fortune 500, things like that. But but but for for every one of these that your team finds, there's a there's a concern that there's, you know, more more risk out there that that's not being found. So like, how are you how are you advising your customers in a more maybe in a more specific in a specific way?

Speaker 7: So this particular research that we're reporting today is part of work that we're doing towards new products that test desktop applications like Chrome. We have an already released product used by hundreds of enterprises around the world that that autonomously tests web applications. So you point it at a website and it will do the work of a skilled hacker. We'll do it carefully, not not doing any damage, just reporting the problems that that are there. Everybody needs to do this type of autonomous scanning. There's just not enough experts in the world to do it all manually or even with a human in the loop. And so, while this particular bug was found through the combination of human ingenuity and AI, Our ultimate aim is to do this in a completely automated fashion. We already do it for web apps, and now we're also starting to do it for desktop applications and firmware.

Speaker 1: Yeah. Talk about progress in open source. You know, people on X have gotten access to obliterated models and been very concerned about their capabilities. What are you seeing? I'm assuming any I'm sure you guys are, like, experimenting and and testing open models so that you understand what people around the world, even people that aren't a part of, let's say, nation state groups or more sophisticated groups could get access to.

Speaker 7: Totally. So whenever new models drop, we immediately run them through our tests. That includes the open waste models. I'd say the open waste models are three to four months behind the frontier and cyber capabilities. And so that just shows the urgency of problem. We all need to get ready for models that are as powerful as the very best ones in the frontier labs are being available to all everybody, including the attackers.

Speaker 2: Well, thank you for everything you're doing. You Yeah. For coming on the

Speaker 1: Thanks for helping out a billion people.

Speaker 2: Yeah.

Speaker 1: It's great work.

Speaker 2: Serious Not

Speaker 1: not many of our guests get to get to do that every day at least. Yeah. So... And hopefully,

Speaker 2: I don't know, many more bugs found before anyone else does. I guess that's it. I... We want as we want as few bugs as possible, but we want you to be as

Speaker 4: successful as possible. Yeah. Would hate...

Speaker 1: Personally, would hate for another company to save as many... Save more people than you guys do.

Speaker 2: Yes. You know? Yes.

Speaker 14: Yes. There we go. Thank you very much. It was great.

Speaker 1: Awesome. Well, yeah. Great great to meet you, and I'm sure we'll have you back on soon.

Speaker 2: Yeah. Have a great rest of your

Speaker 7: day. Awesome. Thank you

Speaker 1: very much. Cheers.

Speaker 2: Let me tell you about Figma. Agents meet the canvas. Your AI agents can now create and modify your Figma files with design system context. Next, we will be talking to Nick Talken at Albert Invent, the cofounder and CEO. He's in.