Efforts to secure generative AI systems are increasingly clashing with a key limitation: many of the most serious risks cannot be regulated or filtered away. New research suggests that existing ...