Defenses — Guardrails & Constitutional AI

Every attack in this track meets the same answer: defense-in-depth, because no single control holds. Safety gets baked into the weights with RLHF and Constitutional AI, wrapped at runtime with input/output guardrails, and double-checked by separate moderation classifiers like Llama Guard and Lakera. Here is how each layer actually works, how they stack, and the honest part — it is an arms race with a real over-refusal cost.