Type: url
Original: https://simonwillison.net/2025/Jun/13/prompt-injection-design-patterns/
Status: generated
Summary
A 13 June 2025 blog post reviewing a multi-institution paper that proposes six design patterns for constraining LLM agents against prompt injection. The post summarizes each pattern — action-selector, plan-then-execute, LLM map-reduce, dual LLM, code-then-execute and context-minimization — endorses the paper’s premise that general-purpose agents cannot currently offer reliable safety guarantees, and notes that one of the six is the reviewer’s own earlier proposal.