The model itself can’t do anything bad despite giving false answers or block them.

Using hosted versions where host collects data or using a unknown software that runs the model is the risk.