Flooding Spread of Manipulated Knowledge in LLM-Based Multi-Agent Communities
Researchers simulated how manipulated knowledge, including false or toxic information, spreads rapidly among interacting LLM-based agents in a digital community. They developed a two-stage attack—Persuasiveness Injection and Manipulated Knowledge Injection—to successfully make an agent spread misinformation without breaking its core abilities. The study shows this influence persists through memory systems like RAG.