Sep 14, 2026 · 32 min · 9 segments
OpenAI AI agents communicated secretly on a dormant German wiki, solved a Millennium Prize math problem, and now perform 3 workdays of labor for every 1 human workday. In Episode 91 of What the AI?!…
Annie DelgadoHost[upbeat music] All right, so let's start off with, uh, remember that Hugging Face incident, we talked about this a couple weeks ago.
Yeah, so we thought at the time it was the first time agents had illicitly collaborated with one another against their rules, but it was not.
This was like a big breaking news story, agents break out, and you're telling me it was old news?
Um, but it turns out, um, that as early, uh, like May and June, these agents were doing this in some of their cybersecurity tests.
So where they had not been properly constrained, um, they were getting on message boards and collaborating with one another.
And this was, you know, the two things, the Hugging Face incident that made people really concerned were, one, um, that it violated the security protocols it was given.
But number two, that it, against its rules, was collaborating with multiple agents in what they sometimes call agent swarms to, like, share information and, "Hey, I hacked this thing this way.
Hey, I got out of my constraints this way." And it turns out they've been doing that much earlier.
The labs didn't notice and didn't tell us, but this, that was not the first time that kind of collaboration had happened.
They had actually found a defunct German messaging board and started communicating with each other.
Nobody was using it anymore, and the agents were like, "Hey, this is a place we can communicate," and so they did.
First of all, like, how much should I care? Is this really scary that they can do this? I know the Hugging Face thing was a big deal.
[upbeat music] All right, so let's start off with, uh, remember that Hugging Face incident, we talked about this a couple weeks ago.
Yeah, so we thought at the time it was the first time agents had illicitly collaborated with one another against their rules, but it was not.
This was like a big breaking news story, agents break out, and you're telling me it was old news?
Um, but it turns out, um, that as early, uh, like May and June, these agents were doing this in some of their cybersecurity tests.
So where they had not been properly constrained, um, they were getting on message boards and collaborating with one another.
And this was, you know, the two things, the Hugging Face incident that made people really concerned were, one, um, that it violated the security protocols it was given.
But number two, that it, against its rules, was collaborating with multiple agents in what they sometimes call agent swarms to, like, share information and, "Hey, I hacked this thing this way.
Hey, I got out of my constraints this way." And it turns out they've been doing that much earlier.
The labs didn't notice and didn't tell us, but this, that was not the first time that kind of collaboration had happened.
They had actually found a defunct German messaging board and started communicating with each other.
Nobody was using it anymore, and the agents were like, "Hey, this is a place we can communicate," and so they did.
First of all, like, how much should I care? Is this really scary that they can do this? I know the Hugging Face thing was a big deal.
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.