I wonder if this is just an example of Goodhart's law. How did they measure performance of those models? I would imagine they tried measuring against known cases of forms misuse, aka those forms that contained 'password' field.