Researchers found that persistent conversations could push major AI chatbots past their biological safety barriers, as increasingly capable models acquire knowledge that worries biosecurity experts.