Garrison Lovely

speaker

631 appearances 3 recordings 1 series first heard Jul 2026 last heard 15 Sep

Garrison Lovely’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

Recordings per month over the last 12 months — 3 in all, peaking in Sep 2026 with 2.

Appearances

newest first · ▶ plays the moment

Breaking Points with Krystal and Saagar · 7/23/26: War Historically Unpopular, AI Hack Meltdown, Old People Debate · 24 Jul 2026 podcast

And it was a test, but if they had access to real-world infrastructure that they could control and shut off,

the blackmail might actually have some punch.

Yeah, yeah.

So basically, the OpenAI models were operating in this environment where they're not supposed to have access to the outside world, the internet, anything like that.

The models were trying to answer this evaluation.

The questions were very hard.

And they're like, well, these questions are hard.

Maybe it'd be easier to just get the answers to the questions.

So maybe to do that, I'll get access to the internet.

So it pokes around in its environment, finds a little chink in the armor,

Then it gets out and starts jumping around to different servers within OpenAI until it finds one that has access to the Internet.

Then it uses that to start looking for what might have the answers to this evaluation.

uh finds hugging face as a candidate and then finds exploits and hugging faces code base that nobody knew existed um and then uses those exploits and like stolen security credentials and like all these classic hacker moves to get access to hugging face and according to opening i successfully found the answer set um

and did this all on its own and without OpenAI noticing for maybe a week or more.

This is like reading between the lines from the Hugging Face report, OpenAI's report.

We don't have a super clear timeline and OpenAI was vague about this because it looks pretty bad if your models both autonomously hack another company, which is a crime if a human did it, but can't be a crime because the models can't have intent legally.

But if a human did this, it would be criminal.

Um, and then you have no idea it's happening until the other company figures it out, reports it.

And then maybe you see the report and you're like, Oh shit, that sounds like us.

And you publish it.

Showing 501–520 of 631 · page 26 of 32