CyberWire Daily · Software Engineering & Infra · November 2026
Asked what surprised him while building Sage, his open-source research harness for AI-assisted offensive security, Van Tuyl says the scaffolding he added around the model initially made it perform worse.
…you as the human have more experience and know that what AI is wanting to do sounds good on paper, but maybe isn't the best OPSEC or you know that that's not actually going to work for a reason. So you have the ability to kind of respond to that and like steer it a different way kind of thing.
What has the development journey been like for you here? Have there been any particular surprises or insights along the way?
I have a bunch of insights along the way, actually. I considered writing a blog post for it, but never made my way to it. One of the things I learned about creating an AI agent harness is that it's actually possible to make the model dumber. Well, I shouldn't say the model dumber. I should say constrain the model so that it can actually work. So some of my first months working on Sage, I was adding these tools and these capabilities that I wanted it to do. And basically I was kind of hamstringing the model from doing, it would have performed better without Sage to begin with. I think it's an opposite side of that now, but I learned…
You've made Sage an open source research project. Why was that important to you, that choice?
We obviously really believe in demystifying adversary tradecraft and like making the things that look like math. Magic known and understandable to everybody in both offensive and defensive security. Namely, so that way both sides of the house and cybersecurity can understand how things work and how…
You know, there's an obvious dual use dimension here. How do you address that concern, you know, both for yourself and other folks who want to take a look at this?