“A method has been developed to forecast potential risks in language models across orders of magnitude more queries than are tested during evaluation.”