Could 'banned topics' be the secret to stopping AI hackers?