Independent safety evaluations are an effective way to prevent AI control failures