This document proposes a new black-box adversarial attack method called Prior Convictions that utilizes gradient priors and online convex optimization techniques. It formulates black-box attack as an online learning problem and estimates gradients using a sphere sampling estimator within a bandit convex optimization framework. The algorithm iteratively improves its gradient estimation by incorporating temporal and spatial gradient priors. Experiments show the method generates adversarial examples for black-box models more effectively than prior gradient-free approaches.