Adversarial Alignment for LLMs Requires Simpler, Reproducible, and More Measurable Objectives