An Adversarial Perspective on Machine Unlearning for AI Safety