The Dual LLM pattern for building AI assistants that can resist prompt injection | Hacker News