Can AI avoid getting into a dangerous situation?