RV Russel Van Tuyl On CyberWire Daily

“Everybody wants AI agents to work in their production environment to find and exploit attack paths at machine speed so that way they can resolve them. I think that is a utopia in the sense that it sounds great until production networks start going down because we are trying to reach that goal.”

CyberWire Daily · Software Engineering & Infra · November 2026

“Everybody wants AI agents to work in their production environment to find and exploit attack paths at machine speed so that way they can resolve them. I think that is a utopia in the sense that it sounds great until production networks start going down because we are trying to reach that goal.” — Russel Van Tuyl, CyberWire Daily

Context

Asked where autonomous AI agents add the most value in offensive security, Russel Van Tuyl says real attackers can let agent swarms loose because they accept the risk. Authorized testers can't take down a client's production network to prove a point.

Transcript

CyberWire Daily Around 13:52 into the episode
Speaker 4

…because we're still trying to work through this research phase of making sure like it's actually reliable. You know, if you pay attention to the news, AI can sometimes do a lot of things that it wasn't intended to do. So, I'm really focused on making sure I have a good control on that part of it.

Dave Bittner

There's a lot of enthusiasm around these autonomous AI agents these days. In offensive security, where do you see that autonomy delivering the most value?

Speaker 4

…single person. If you're looking at cybersecurity, generally speaking, it can obviously find vulnerabilities a lot quicker. And then you can use more of the offensive security tools to like demonstrate if there's actual value or an actual exploit there to kind of keep pace with that. But it really depends on where the AI is at in that part. When we talk about offensive security work, I would say everybody wants AI agents to work in their production environment to find and exploit attack paths at machine speed so that way they can resolve them. I think that is a utopia in the sense that it sounds great until production networks start going down because we are trying to reach that goal.

Dave Bittner

I know one of the challenges you've identified is that the AI models can confidently report that something worked when it didn't. What does that look like in practice? And how do you put guardrails on that?

Speaker 4

Yeah, I spent a lot of time thinking about that in Sage, like how to make sure the model isn't lying to you or like confidently thinking that something is there. One of the ways that I do that in Sage is there's this thing called the engagement state ledger. And it kind of tracks like each of the…

Dave Bittner

What happens if something goes wrong, say halfway through an operation? How does the system recognize failure and either recover or hand control back to a human?

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

More from CyberWire Daily