may never be secure

Is there hope? [...] Probably not.

I have to read the cited paper because this seems like bullshit. Similarly to "Garbage In, Garbage Out" (referring to the quality of the training data), how about "Secrets In, Secrets Out"?

Don't include passwords.txt in the training data, how hard can that be?