“It's not enough to spend one token more than the attacker. It's not even enough necessarily to spend 10x more than the attacker. You kind of want to know exactly which models the attacker is using and where they're looking. … If Mythos and DeepSeek are finding the exact same things, then you should just be running DeepSeek very aggressively and very expensively over most of your application.”
Open Source Startup Podcast · Hands-On Engineering Podcasts · November 2026
Context
Calling AI vulnerability hunting a 'vulnerability slot machine' where enough tokens eventually hit a jackpot, he pushes back on the frontier labs' framing that defenders only need the best model. Defence, in his view, means matching whatever attackers can cheaply run.
Transcript
…to see way more of that. And security teams will actually get genuinely excited about, wow, like, hey, like we wanted to train the developers. Now we can just train the models to like do the things the way we know is going to actually reduce risk to the business. So that's actually really cool.
…type of vendor, right? Finding issues for you, do you find yourself starting to even like invest in surface searching, like agentic research on security side for your own products? And how much you thinking that makes sense for you to invest in, given that it's an unbounded exploration space here?
Absolutely. And this is where I want to be really thoughtful because in a lot of ways, what the labs are doing, the non-charitable phrasing would be stunt hacking in the hacking community. It's like, whoa, like, look at this cool vulnerability that I found. But, you know, most companies until a year ago are not paying people to like sit there and find vulnerabilities inside their products. In fact, like most of them get really annoyed when like a Google Project Zero finds a vulnerability inside their product, right? It's like, oh, like just stop, just stop looking. Some people could have their way, like that's what we would actually do. So there's a difference between vulnerability research, which has typically been the domain of just super intelligent people. With a lot of domain expertise, and is now like models are just better vulnerability researchers than the humans themselves. I think a difference between that and security engineering, where security engineering is much more about we know there are going to be vulnerabilities rather than finding them all. What are we doing to actually reduce their impact to like decrease the statistical probability of them appearing to…
That is interesting because you can kind of say, okay, like obviously LLMs are very new. The ability to use tokens to find vulnerabilities is new, but you could have also just hired and like spent a ton of money to find these before too. It was just a human labor problem. Right. It is interesting…
I mean, I think I am learning more from founder friends, right, who are like founding companies today. Some of the things that I'm learning from them are like, you know, the importance of like capturing the traces and like doing analysis on like the traces and the intention behind the developer…
Super interesting. And I'm so happy that we had you on because I think this is one of the most important topics. You're kind of at the like middle of a lot of the security conversations going around agents and AI. So, thank you so much for doing this with us, Isaac.
More from Open Source Startup Podcast
