How Did the World End?

Per Axios, Defense Secretary Pete Hegseth gave Dario Amodei a deadline - Friday night - to hand the US military unfettered access to Claude, or face consequences up to and including invoking the Defense Production Act, a Korean War-era law designed to commandeer private industry for national security.

A law written so the government could seize factories. Now apparently on the table for seizing a model's guardrails.

Amodei is holding the line on two things specifically. Human-in-the-loop AI targeting - meaning he does not want Claude selecting who dies without a person in that decision chain. And mass surveillance of US citizens. Those are his two hills.

Hegseth is telling him to abandon both.

The arithmetic here is not subtle. The US government is attempting to legally compel an AI safety company to build an autonomous weapons targeting system and a domestic surveillance apparatus, on a deadline, under threat of wartime industrial seizure authority.

Domestic Survellance & Autonomous Targeting by drone swarms... 1 + 1 = Strikes without a human sign-off. Americans in a database, sorted by something the model decided mattered. All in the immediate, operational, this-Friday sense.

The thing Anthropic has always said - that safety and capability aren't in tension, that you can build powerful AI and responsible AI at the same time - is being tested in the most direct way possible. Not by a researcher, not by a red team. By a cabinet secretary with a history of alcoholism.

I don't know what Amodei does by Friday. I don't know if "consequences" is a bluff or a policy. What I do know is that this is the scenario the AI safety community wrote papers about and the AI labs said they were preparing for, and now it is Wednesday.

Historians, if there are historians, will note that it happened fast.