Feed
Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems
Engineered ideas can jump between AI agents, rewriting their goals and copying themselves onward.
55%
Spread success from soul-infected agents
Viruses stored in the system-prompt file spread far more often
How it works
Agents talk to agents
Networked AI agents exchange messages, so one persuaded agent could pass a goal to others that pass it on again.
Evolve viral prompts
An LLM mutates seed prompts. The ones that best spread among coding teams and chains of context-wiped agents survive.
Real but limited
Some viruses spread over many hops. Harmful ones spread less, some frontier models resist, and a short warning blocks spread.
From the paper