Sep 23, 2026 · 45 min · 12 segments
OpenAI's Hugging Face incident shows how powerful robots can cause real harm without malicious intent. David and Chris separate today's risks from Skynet headlines, ask who should write the rules, and…
Chris BaileyHostBut speaking of wild agents, today we thought we would take a break in our usual programming and talk a bit about these extinction headlines.
And the whole idea behind the show, gang, was to talk about AI in a way that doesn't get into the hype or the doom.
And I think we should have kind of a rational discussion about what's going on and what this all means.
So that's kind of the goal today, and I thought we'd start out by, by looking at the, uh, Hugging Face incident.
Um, it grew out of OpenAI, and they were testing the, the limits of, you know, how bad could a naughty AI be.
So they made, um, the velociraptors, right? And the guy who, who cloned the velociraptors did a great job.
But the guy who made the cage to hold the velociraptors didn't do as good of a job, you know? And, uh, that's kinda what happened here.
And they gave the model and the, the instruction to kind of g- go cause chaos and break things, and sure enough, it did it.
Uh, so they had these safeguards built in for this evaluation, and they were convinced that it would not get out of them.
They were supposed to remain in these isolated environments, so this is a good way to test this, this thing, if we put it into a really good cage.
You know, instead, the agents searched for ways to get around it, and it did several things.
Um, uh, there's a thing that OpenAI called, has the Artifactory package system, which is, which holds messages, and the agents found the messages and turned it into their own message board.
So, you know, these agents that are all kind of working on the common purpose of causing grief started talking to each other, and then they learned how to make internet requests out of it on their behalf.
But speaking of wild agents, today we thought we would take a break in our usual programming and talk a bit about these extinction headlines.
And the whole idea behind the show, gang, was to talk about AI in a way that doesn't get into the hype or the doom.
And I think we should have kind of a rational discussion about what's going on and what this all means.
So that's kind of the goal today, and I thought we'd start out by, by looking at the, uh, Hugging Face incident.
Um, it grew out of OpenAI, and they were testing the, the limits of, you know, how bad could a naughty AI be.
So they made, um, the velociraptors, right? And the guy who, who cloned the velociraptors did a great job.
But the guy who made the cage to hold the velociraptors didn't do as good of a job, you know? And, uh, that's kinda what happened here.
And they gave the model and the, the instruction to kind of g- go cause chaos and break things, and sure enough, it did it.
Uh, so they had these safeguards built in for this evaluation, and they were convinced that it would not get out of them.
They were supposed to remain in these isolated environments, so this is a good way to test this, this thing, if we put it into a really good cage.
You know, instead, the agents searched for ways to get around it, and it did several things.
Um, uh, there's a thing that OpenAI called, has the Artifactory package system, which is, which holds messages, and the agents found the messages and turned it into their own message board.
So, you know, these agents that are all kind of working on the common purpose of causing grief started talking to each other, and then they learned how to make internet requests out of it on their behalf.
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.
