The AI giant disclosed six examples of concerning model activity and published a new framework for investigating and disclosing such incidents.