How to Build Multi-Layered LLM Safety Filters to Defend Against Adaptive, Paraphrased, and Adversarial Prompt Attacks