Ask HN: Deep learning attack vectors

Since the possibility of living in a world full of AIs based on deep learning is almost a fact now, what would be the weak points that could be explored to throw an AI off the tracks? Is deviating from the learning data the best approach?