Eight Ways to Break a Multi-Agent System

What I learned building a deliberately vulnerable LangGraph system and attacking it Most prompt-injection writing treats the LLM as a single box: one prompt in, one completion out, one place for the a