Describe how you would build a scalable evaluation pipeline to test the robustness of an LLM against adversarial prompt injection.
