finetune a model to generate huge lists of possible passwords, using a database as a train/test set, success being producing lists that contain real passwords with greater likelihood

could stretch the data some and extend training by rerandomizing the token mapping

-

maybe a solution to all this would be to bind user-specific passwords to user-specific secret keys.  could have the password "houseguest" work for 5 devices or such.  I wonder what other protocol improvements have been proposed.

... router could use its antenna to identify who you are and what you are doing from your biosignals using classification models.  wave of the future.