If I understand correctly, the main problem here is that if a password manager at some point asks you for a password in an online environment, they're subject to coercion. This is especially dangerous if you're using auto-updating code like Javascript in a browser or code on a remote service, because it could get backdoored at any time and you wouldn't notice. Isn't the real problem auto-updating code with access to…
At least in theory you can sandbox an app so that it does not have (unlimited) access to the network.