After AI extinction theory went viral, Anthropic granted external reviewers near employee-level access.

Beating AI News Flash: Former Anthropic researcher Jacob Coxon's resignation post surged to 160 million views in just over three days, pushing AI extinction risk rapidly into the mainstream American public discourse this week. OpenAI Chief Scientist Jakub Pachocki had previously called on peers to coordinate a slowdown if necessary; after Coxon's post went viral, Sam Altman also said internally that he was willing to slow down R&D together with other labs, and OpenAI began pushing for mandatory national AI safety regulation.

Now Anthropic has also taken concrete action. The company has committed to having a third-party safety team permanently on-site, giving them desks, access badges, company computers, and tools and permissions close to those of its internal risk assessment team. They can inspect training processes, safety measures, and incidents, and can independently publish their conclusions externally. Anthropic can redact information involving safety, legal, and customer privacy, but cannot delete conclusions just because they look bad. Anthropic CEO Dario said this approach goes further than any AI company currently does.

Dario's attitude toward "slowing down" has also changed. He said that when the industry discussed pausing AI in 2023, he still thought it was meaningless. Now AI is increasingly involved in developing the next generation of AI, and capability progress is starting to accelerate. He believes that doing safety research alone is no longer enough, and that model capabilities themselves should also slow down.

He also called out OpenAI's Hugging Face失控 incident, warning that at the current pace, in another 6 to 12 months, similar Agent swarms may already have the ability to build botnets covering large-scale internet systems, causing hundreds of billions of dollars in losses. Next, he hopes leading AI companies in the United States and other democracies will unify safety standards and R&D speed, and then try to coordinate the pace of global AI development with China. However, restrictions on advanced chips and model distillation should continue, so that the United States can maintain its lead while slowing down.