Researchers at Anthropic taught AI chat bots how to lie, and they were way too good at it