The company's former safety lead said frontier AI model releases should have "layers of redundancy and careful, time-consuming planning."