Identify harmful capabilities and outputs, create safety policies and evaluations, add layered safeguards, and preserve meaningful human oversight.
