Architectural defenses keep untrusted content away from the component that plans and calls tools. The dual LLM pattern uses a quarantined model that never gets tool access, and CaMeL extends it with explicit control flow, data flow tracking and capability checks. The design patterns paper catalogs six such patterns with their utility tradeoffs.
Read in this order
- ArticleThe Dual LLM pattern for building AI assistants that can resist prompt injection (Simon Willison)Start here: the original dual LLM idea, explained in plain terms.
- VideoAI Security with Guarantees ([un]prompted 2026)Introduces architectural guarantees for agent execution. 26 min.
- Open sourcecamel-prompt-injection (Google Research)CaMeL's reference code, to see capability checks in practice.