Design Patterns for Securing LLM Agents against Prompt Injection

Luca Beurer-Kellner and other authors propose a set of design patterns for building LLM agents that are provably resistant to prompt injection. The paper systematically analyzes these patterns, discusses their trade-offs between practicality and security, and demonstrates their real-world applicability through case studies. The authors point out that the risk of prompt injection is especially prominent when agents gain tool access or handle sensitive information.