Company dinners at Anthropic in its early years kept returning to the same scenario. The staff would be asked to move to the desert, and work would continue inside an electromagnetically shielded government facility. Former employees who remember those conversations describe a Manhattan Project arrangement rather than a joke.
That instinct has since moved from talk toward paperwork. A small group of the company’s earliest employees told an industry acquaintance in recent weeks that they are shopping for land in remote parts of the United States, according to the Wall Street Journal, as somewhere to go if AI slips beyond human control. Nobody has closed a purchase. Nobody involved has spoken on the record.
Credit for the worldview usually goes to Eliezer Yudkowsky, who has spent two decades arguing that building AI could end badly. Reporters trace the network he influenced to the Bay Area effective altruism scene, which supplied the first rosters at both Anthropic and OpenAI. Its members spent years gaming out end-of-the-world scenarios, long before AI reached the top of their threat list.
In public the company takes the opposite posture. It publishes the policies it says govern deployment and brands itself the lab that treats safety as the point. The report suggests part of its own workforce is planning around the chance that those policies do not hold.
