In order to develop robust grasping AI algorithms, it is important to generate objects that become progressively more adversarial and complex with each iteration, as in curriculum learning, in a manner agnostic to the particular visual processing method used by the robot. We present a data generator of voxel-based 3D adversarial object geometries for training grasping algorithms, which operates independently of any particular visual encoding/decoding pipeline. Our novel approach uses genetic algorithms and can generate arbitrarily many samples, bounded only by the user-specified voxel resolution. The success of our approach is demonstrated in terms of the number of adversarial objects generated, how difficult they are to grasp, and how similar they are to more easily graspable objects.