Self-replicating prompts can spread from one agent or user to the next through email, shared documents, or agent-to-agent messages. These attacks show why a single injection in a connected ecosystem can reach far beyond its first victim.
Read in this order
- ArticleHere Comes the AI Worm (Morris II project site)Start here: the project site for the first AI worm, with a demo.
- PaperHere Comes The AI Worm: Unleashing Zero-click Worms that Target GenAI-Powered ApplicationsThe full paper on zero-click worms against assistants that read email.
- PaperPrompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent SystemsHow an injection spreads between agents inside one multi-agent system.