00:09:23 · Rebecca Balan, Tomer Schwartz
[00:09:23] S1: Yeah. I mean, it's great for defenders. It's also great for attackers. Right? And that's why we saw such a push to get this bug fixed because attackers are using AI to send out mass attacks and to 100x their own capabilities. Can you tell me a little bit more about Atlas? Atlas is your harness. Using, you said, OpenAI, Anthropic models. Are you using latest site? Are you using Mythos in that? Do you have access to that yet? Which models are you using and playing with? [00:09:51] S2: We constantly change it. With the pace in which AI models get better and better, we've to update this all the time, and there's also various different steps where we find that certain models are better at as opposed to others. [00:10:06] S1: Can you give me an example?
00:10:06 · Rebecca Balan, Tomer Schwartz
[00:10:06] S1: Can you give me an example? [00:10:08] S2: So let's talk about binary exploitation versus code comprehension. Right? These are very different tasks. If I take a very large code base and I'm looking for, hey, what kind of what kind of interactions do do we see here between components and what do they assume about how other components are using them? Maybe maybe there's one model that's really good about processing large chunks of data and compressing it into a whole bunch of insights that we can use later. But then if the the only thing we need is to to get to a certain part of the code base and figure out how to trigger this remotely, this is a completely different task. So we are trying to evaluate different models all the time. We we actually have a publicly facing benchmark that we we release. Fortunately, we also have access to unreleased models as well with the the labs that we're working with. So what we're trying to do is to just sort of do a best of breed kind of combination and see what's what's the best solution for this problem. [00:11:05] S1: How much of the mix is open models? Because I know a lot of people I've spoken to, including, you know, Clint along from Hugging Face, say that open models are really great cybersecurity because you don't run into any of the guardrails when you're red teaming or searching for vulnerabilities that you might with some of the frontier models. Is that what you're seeing as well?
00:11:05 · Rebecca Balan, Tomer Schwartz
[00:11:05] S1: How much of the mix is open models? Because I know a lot of people I've spoken to, including, you know, Clint along from Hugging Face, say that open models are really great cybersecurity because you don't run into any of the guardrails when you're red teaming or searching for vulnerabilities that you might with some of the frontier models. Is that what you're seeing as well? [00:11:22] S2: Well, open weight models are good, and they're getting better and better. But honestly, we're fortunate to be a part of Google right now, so we have access to pretty good models at unique pricing, I would say. But other than that, I think that that when you're really thinking about scaling research in in massive ways, maybe some of the startup companies are listening to us right now in podcast or or thinking about this too, we did a lot of different experiments with open weight models, and we do find them to be extremely effective. We had a couple of iterations where we did use them. It's just that the current model that we're using, the current harness that we're using is not using any open weight models. No. [00:12:04] S1: Not using any open weight models. Interesting. [00:12:06] S2: Not right now, but that situation could change in a month or two. It's kind of fun to see how quickly the the AI world is is moving.
00:12:06 · Tomer Schwartz, Rebecca Balan
[00:12:06] S2: Not right now, but that situation could change in a month or two. It's kind of fun to see how quickly the the AI world is is moving. [00:12:16] S1: Yeah. Absolutely. Now, you know, because it's moving so quickly, which was part of the problem with, you know, rushing to get the fix out once the word was out that there was this bug, Did you learn anything from this? Did Wiz learn anything from this in terms... Because I know you had a schedule when you wanted to release this publicly. Have you learned anything in terms of how to handle doing this quietly and quickly? What was it like inside? Did it ruin your weekend? What happened?
00:12:40 · Tomer Schwartz
[00:12:40] S2: There's always learnings out of this, and coordinated disclosure is something that is fundamentally flawed. We want to bring more and more people into the inner circle to make sure that defenders are aware of the the vulnerability and they have time to prepare. But then on the other side, we can't really announce it to the world without adversaries know... Knowing about this. And this is not a model that we invented, obviously. This this is a problem in the cybersecurity industry for decades. The question of how to do it well and how to do this in a very timely process, I think, is is right now a big question that the industry needs to tackle with the the the imminent vulnpocalypse that we're seeing right now. So I think it was a a a pretty fast turnaround time compared to to all the things that needed to happen. From detection to fix, I think it was roughly two months, which is pretty amazing given the fact that this this was a a very unique architecture flaw within PHP that was was lurking there for twelve years. Right? It was it was very diff... Difficult for us to come up with a solution that is reliable enough, that covers every single aspect of this unique vulnerability, that is backwards compatible with some of the extensions that PHP has. There were a lot of different constraints that we had to operate in. And given the fact that we couldn't really bring in a lot of people into that inner circle, I think we managed to do this, again, comparably pretty quickly. The other side of this is what happened towards the end of that cycle where we realized, well, this is something that can take the Internet down quite literally. So we got to work with the cloud providers, with the the web application firewall providers, with, you know, lot of different vendors that are associated with this that need to be aware so they can make sure that they release on time, they have the the detection rules out, that we have we have the full set of visibility and and potential... Not full defense, but partial defense against what's what's gonna happen. And once you realize how big this industry is, you realize that, you know, maybe that inner circle is all all of sudden hundreds of people. And maybe they're talking to some some of their colleagues because they have to pull them in to be able to operate. So I don't I don't blame anyone for this leaking.
00:15:00 · Tomer Schwartz, S0, Rebecca Balan
It's just... It's it's it's a function of how coordinated disclosure is happening. And luckily, we were pretty much ready to that situation. We had sort of a plan b that we we could have operated bay... Based off of. And I think that we made the right decisions in in in every given point in time. [00:15:18] S0: Every company today is looking for an edge. Better insights, faster decisions, greater resilience. But as AI drives new growth, it also creates new demands on energy and infrastructure. Schneider Electric is connecting energy, automation, and digital technologies to help organizations make smarter decisions and operate more efficiently. Because the future won't be shaped by AI alone, it will be shaped by energy and industrial intelligence. Schneider Electric, advancing energy tech. [00:15:45] S1: So just to like take a step back for coordinated disclosure for anyone who isn't familiar with this, researchers will privately notify software maintainers, give them time to prepare a fix before the vulnerability is made public. Right? And so PHP Foundation would have given big companies and hosts a bit of a head start. Maybe some people had already, you know, implemented a fix or started to before it became public. I don't know.