Adversarial manipulation of training data to degrade model behavior while maintaining normal performance on standard metrics.