"AI is great at following patterns", sure but if some attackers can "inject attack patterns" at a high rate in your training set (check "dataset poisoning") then it's going to be an issue...