“Misspecification of reward functions is a significant concern in AI safety circles, as an AI might achieve a stated objective in unintended and harmful ways.”