How to Stop Prompt Injection in AI Agents: Enforced Guardrails, Not Prompts