Coming AI regulation may not protect us from dangerous AI