If you're looking for a purely topos-theoretic model, I think you don't need to go through set theory (even though the end result may end up being basically equivalent). Look at the topos of continuous actions of the pro-completion of the integers, which is to say, the category of sets equipped with an automorphism all of whose orbits are finite. Here we have an N-indexed family of objects (one orbit of each cardinality) which are all inhabited, but whose product is empty -- hence the NNO is not internally projective.

Have you read P. Freyd's paper "The Axiom of Choice"?