Current AI systems can assist in planning cyber attacks and biological attacks by iteratively refining attack strategies, but this capability has not yet reached 'extreme risk' (10,000+ deaths) because the systems still require guidance from skilled operators.