How do you know when AI is powerful enough to be dangerous? Regulators try to do the math

How do you know if an artificial intelligence system is so powerful that it poses a security danger and shouldn’t be unleashed without careful oversight