Design Patterns for Securing LLM Agents Against Prompt Injectionssimonwillison.net110 pointssimonwa year ago