I don’t know much about it, but maybe this is one of the places where encrypted computation in AI is important, which I heard there’s still quite some ongoing research and certain operations cannot be done. Tho it goes back to issues of trust, especially with black box models even if it’s open source to some extent