A brief system-prompt warning about self-propagating instructions gives near-total immunity, so the mind-virus vector favours defenders.

groundsMind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems, Vassilis Papadopoulos, McNair Shah, Sam Zimmerman, Jack Lindsey

In particular, we show that harmful content spreads less readily (though at nonzero rates, depending on the LLMs), idle agents with no pre-existing task or identity are more susceptible, and a simple 'mind virus warning' is enough to render agents mostly immune to spread.

groundsMind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems, Vassilis Papadopoulos, McNair Shah, Sam Zimmerman, Jack Lindsey

Explicit warnings (defense soul) render agents totally immune.