Neural networks: train mlp (input x, output y) with a black box that labels any x-y pair good / bad

The standard MLP (xay map) is trained with a set of x-y data points.

My question: What happens if there is no train data, the only supervisor leaves any x-y pair with 0 or 1. The goal is that the x-y pair generated by the MLP is always labeled 0. How to train the MLP?