Transferable End-to-End Optimization for Indirect Long-Term Memory Poisoning in LLM Agents
The episode discusses the paper "Transferable End-to-End Optimization for Indirect Long-Term Memory Poisoning in LLM Agents." Hosts analyze how attackers use untrusted external content to poison an AI agent's memory across a three-stage lifecycle: writing, retrieval, and utilization. They conclude that the solution, P IPE P OISON, improves attack success by optimizing the entire pipeline simultaneously.