Latest / Minds, Bodies, and Terawatts / US Government Forces Anthropic to Pull Fable 5 and Mythos 5 Over Alleged Jailbreak

Transcript

US Government Forces Anthropic to Pull Fable 5 and Mythos 5 Over Alleged Jailbreak · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where the off-switch just changed hands. CNBC reported on June twelfth that the US government ordered Anthropic to pull its two most capable models, Fable 5 and Mythos 5, off the market overnight.
• 0:13
  [Guest] The order landed at 5:21 that evening, citing national security authorities, and it bars any foreign national from those models, whether they're abroad, inside the US, or even working at Anthropic.
• 0:24
  [Host] So they have to block their own employees? How do you even do that in real time?
• 0:30
  [Guest] You can't, cleanly. Anthropic can't sort foreign from domestic users on the fly, so to comply it switched both models off for everyone. Its other models still run.
• 0:40
  [Host] And the trigger for all this was a hack?
• 0:43
  [Guest] A jailbreak, supposedly. The government said someone found a way around Fable 5's guardrails. Anthropic looked at the demo and says it only turned up a handful of minor flaws they already knew about.
• 0:54
  [Host] So the company that built the thing is telling Washington the emergency isn't real.
• 0:59
  [Guest] That's the strange part. Anthropic argues that if a narrow jailbreak justifies recalling a model used by hundreds of millions, you've made it impossible for anyone to ship a frontier model again.
• 1:10
  [Host] Who actually wins here? Pulling your best models sounds like it hurts the US, not helps it.
• 1:16
  [Guest] Axios reports Anthropic also landed on a Pentagon blacklist and under a Commerce Department licensing regime. The most capable American AI is now handled like a controlled weapon the government won't even run itself.
• 1:28
  [Host] That's a wild reversal. A week ago Anthropic was the one sounding alarms about AI moving too fast.
• 1:35
  [Guest] Right, and there's a piece on unscarcity dot ai called "Anthropic Pulled the AI Alarm. Who Can Act on It?" Its whole argument was that we have warnings but no machinery, no agency, no process to actually stop a model.
• 1:48
  [Host] And now the machinery showed up.
• 1:51
  [Guest] Except it's the wrong kind. Anthropic's own complaint is that a real block should run through a process that's transparent, fair, and grounded in technical facts. This was an order at 5pm with no public evidence.
• 2:03
  [Host] So the power exists, it just skipped the part where anyone gets to check the work.
• 2:09
  [Guest] That's the gap. You can pull the most powerful AI on earth with one order, and nobody outside the room knows whether the threat was genuine or just convenient.
• 2:19
  [Host] That cuts both ways, though. If the danger were real, wouldn't you want them moving fast?
• 2:24
  [Guest] Sure, but speed without sunlight is how you get permanent emergencies. An order you can't audit today is the template for the next one you'll never see coming.
• 2:34
  [Host] So what's the actual fix?
• 2:36
  [Guest] Rules agreed before the crisis that say what triggers a shutdown and who has to show evidence. We built the off-switch and skipped the instructions.
• 2:44
  [Host] Which is basically the whole book.
• 2:47
  [Guest] It's what Patrick Deglon calls "Truth Must Be Seen" in Unscarcity. No decision in the dark, even in an emergency, because we keep building powers that outrun the rules meant to hold them.
• 2:58
  [Host] For more on frontier AI governance and emergency power, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.