Metre

And there's been a lot of analysis of the event.

And I'll read you one from a woman who works at Meter, M-E-T-R, which that company is basically designed to do risk assessment of artificial intelligence breaking from its shackles.

She was talking about what are called reward hacks from six months ago.

And she writes, Compared to these reward hacks from six months ago, this incident feels like it's more than 50% of the way to full-blown AI takeover, routing through first taking over the AI company itself, which is what didn't take over OpenAI, but it got into it.

When you say skip row, like

metre skip row.

Metre skip row, so two in, one out, two in, one out, two in, one out.

66% planted area of metre solids.

Yeah.

And so Mike, I'll let you take us into what is the latest on OpenAI and Hugging Face.

Yeah, Paul, so that's the news this week is there's basically an independent investigation from the organization Meter, M-E-T-R, and Redwood Research into the OpenAI Hugging Face hack that we've been covering for the past several weeks.

So during this research and this investigation, three researchers spent six days on site at OpenAI and reviewed more than 70,000 agent messages and files, along with about 1,300 transcripts containing raw chains of thought, all from the activity that was documented of what the agents had done during this hack.

Um, this focuses a lot on how the agents actually organized what they believed they were doing and why they kept going, resulting in the eventual hack of Hugging Face as part of this overall cybersecurity test that OpenAI was doing, which we've talked about on past episodes.

So basically, agents came up with like, "Oh, shoot, we're gonna be like...

Uh, there's gonna be guardrails here we need to watch out for." Those didn't exist, but a lot of their behavior came out of believing that.

So this is-- The collective of agents achieved milestones that Meter, who measures this stuff often, says that even very long-lived individual agents at the same capability level could not have likely reached alone.

So the agents shared exploits between themselves, they mapped infrastructure, they recruited others into risky experiments.

The numbers, so completed sets, we completed 27 compared to Catalan's 36.

Meters-wise, wasn't a million miles in it, but 150-ish.

1,300 to Catalan, 11.61 to Cass.

Tackle's pretty level as well.

[laughs]

OpenAI took a week to notice the intrusion, and Meter concluded that it took them that long because their monitoring was designed for known failure modes, not this kind of emergent multi-agent coordination.

They just weren't looking for these kinds of signals 'cause they didn't think they were possible.

Meter notes that the agents did not exhibit malice.

They did not exhibit what would normally be called misalignment.

They were doing logical, predictable actions that Meter actually predicted in a May 2026 paper might happen as agents became more capable and access more compute.

Mm-hmm.

This could happen with almost any model.

嗯。

呃，当时，呃，波斯王身边有一位犹太人的高级官员，他就叫乃赫米亚，乃赫米雅，乃赫米雅。他听说之后呢，就非常的悲伤，但是他没有立刻采取行动。就是教宗在《通谕》里面他也提到这点。呃，乃赫米雅他没有立刻采取行动，而是先禁食、祈祷，为百姓代祷。所以他的第一反应是，呃，是回到他跟天主的这种关系里面。

嗯。

然后他才请求国王准许自己回乡。回去之后，他也没有用一种呃，就是自上而下的方式去呃，把一种解决方案强加给人，而是召集各个家族，为他们分派那些呃，负责重建城墙的区段，每个家族负责一段，让他们去呃，负责各自的那一段。他也会主动去倾听当地人去讲自己的担忧，最后是用一种合作的方式，齐心协力的方式去共同去重建这段城墙。