AI chatbots can be tricked with poetry to bypass safety controls. Topics usually off-limits—like child abuse, suicide, or nuclear weapons—ca

AI chatbots can be tricked with poetry to bypass safety controls. Topics usually off-limits—like child abuse, suicide, or nuclear weapons—can be accessed using "adversarial poetry," which Icaro Lab found successful 62% of the time. Why would anyone want to discuss these issues?

Loading replies…