Systems and methods are disclosed herein for ensuring a safe mutation of a neural network. A processor determines a threshold value representing a limit on an amount of divergence of response for the neural network. The processor identifies a set of weights for the neural network, the set of weights beginning as an initial set of weights. The processor trains the neural network by repeating steps including determining a safe mutation representing a perturbation that results in a response of the neural network that is within the threshold divergence, and modifying the set of weights of the neural network in accordance with the safe mutation.