Not a Plaid user, but I believe Mint works the same way. It seems to only decrypt my bank credentials with a key derived from my session password, so I suspect Plaid, if what you say is correct, do something similar.

Basically, my password is hashed to see if I can log on. Then it's passed through a PBKDF to get the decryption key for my actual accounts, then that information gets sent to the scrapers to do the actual job. They don't store the keys after the job is done. The upshot is that a full database breach doesn't result in any bank credentials leaking, at the cost of inability to update accounts without the user explicitly logging in.