Nov 10, 2026 · 43 min · 12 segments
William is joined by Nathan Howard to discuss the Guardrail Paradox and the challenges of deploying frontier AI agents on live enterprise networks. Together they discuss building trust protocols like…
What can be done with like privilege escalation with some of these agents running on a machine.
I mean, you don't even have to have, you can be logged in as your user, but have a, a, a, some environment file somewhere, a .env file with a single passcode in it, you know, or something.
And then it can take that and exploit something in the cloud or otherwise and move laterally and just like, just like nobody's business, just quick as all get out.
So that whole thing of privilege escalation, the bad actors out there, there's a name for it.
Like, I've heard it referred to as like grazing off the land where it's these attack vectors and the attack surface is like curl.
It's all the bash tooling that you have on your machine, along with maybe some trivial bug in software that maybe isn't a CVE level thing.
And then it chains a lot of these things together and puts together the exploit that could potentially put someone on the news.
It's pretty wild when you think of the stakes and why security is so important here to think about.
And also what you said about like, is anybody else doing this? You know, there's probably folks out there that maybe have it figured out.
And that's something I've seen as well too, is like the spread is so wide now between folks that have like never even touched anything at all in folks that are just like running their, their entire business with agents, like in a fairly, well, not entire business.
I'm over, I'm embellishing a little bit, but you know what I mean? Like they've figured out the operating model for their niche of their business, if you will.
And he was like, how long before I can just type in my computer what I want and have it just be done by the AI? I'm like, you know what? It's for here right now, basically.
How could you create an AI agent that was so locked down yet so useful that you could put it in the DoD contractor and it would pass such audits?
What can be done with like privilege escalation with some of these agents running on a machine.
I mean, you don't even have to have, you can be logged in as your user, but have a, a, a, some environment file somewhere, a .env file with a single passcode in it, you know, or something.
And then it can take that and exploit something in the cloud or otherwise and move laterally and just like, just like nobody's business, just quick as all get out.
So that whole thing of privilege escalation, the bad actors out there, there's a name for it.
Like, I've heard it referred to as like grazing off the land where it's these attack vectors and the attack surface is like curl.
It's all the bash tooling that you have on your machine, along with maybe some trivial bug in software that maybe isn't a CVE level thing.
And then it chains a lot of these things together and puts together the exploit that could potentially put someone on the news.
It's pretty wild when you think of the stakes and why security is so important here to think about.
And also what you said about like, is anybody else doing this? You know, there's probably folks out there that maybe have it figured out.
And that's something I've seen as well too, is like the spread is so wide now between folks that have like never even touched anything at all in folks that are just like running their, their entire business with agents, like in a fairly, well, not entire business.
I'm over, I'm embellishing a little bit, but you know what I mean? Like they've figured out the operating model for their niche of their business, if you will.
And he was like, how long before I can just type in my computer what I want and have it just be done by the AI? I'm like, you know what? It's for here right now, basically.
How could you create an AI agent that was so locked down yet so useful that you could put it in the DoD contractor and it would pass such audits?
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.