Challenges of maintaining AI safety