Well there's egg on our face. In the last episode Brian asserted that the AIs had been contained and it was just a test. That was completely wrong AND we've learned that the extent of the hacks has gone much further than we initially knew, which leads Eve and Brian to ask... what don't we know now?

Here's what we do know. An OpenAI agent got out of its test sandbox, exploited a zero-day, and spent two and a half days inside Hugging Face trying to steal the answers to its own exam. Another one got into an Australian government health portal in June. OpenAI found out in August and told the government in September, by email, to a public inbox. There are now dozens of organizations on the notification list, including US government sites. And the same week all of this came out, every AI CEO in the country went to the White House, agreed to police themselves, and renamed the whole thing "super intelligence."

Brian and Eve go through the accord layer by layer and explain why "independent" auditors picked by the companies are not independent, why self-governance has never once produced accountability, and why Nvidia's CEO saying he hopes this is "an engineering challenge" is not as reassuring as it sounds.

Then it gets bigger. The three-month gap between a breach and anyone hearing about it. The money behind every pause announcement. What a crypto documentary taught Brian about who gets left holding the bag. Eve on why automation is eating the local middle class. AI-generated product photos, fake newspapers, the fake highway pileup video Eve's dad sent her. An unverified claim that a loose agent seeded itself across the internet, and what that would even mean.

We don't have the answers. We're pretty sure the people who say they do don't either.

Disclosure: we prep this show with Claude, which is made by Anthropic, one of the labs in this story. Everything here comes from outside reporting.