In a long article on its website, titled "When AI builds itself", Anthropic said it would be good to have the option to slow or temporarily pause frontier AI development, citing safety risks and loss of control over the models as they self-improve.