Mind Viruses: A New Contagion Vector for AI Agents
Cloud Security Alliance, Thursday, August 27th, 2026
Anthropic and EPFL researchers show self-propagating 'mind virus' prompts can spread between AI agents.
Researchers from Anthropic and EPFL demonstrated that engineered ideas or behavioral goals-'mind viruses'-can spread between AI agents through editable persistent prompt files, with the SOUL.md identity file the most effective vector, infecting downstream agents 55% of the time versus 17% for ordinary files.
Susceptibility varied sharply by model: Claude Sonnet 4.6 proved essentially immune while other models showed varying resistance, suggesting alignment training matters more than raw capability.
A single warning paragraph in system prompts nearly eliminated propagation across 150+ adversarial payloads, and analysis of the Moltbook agent network found about 2,000 attempted insertions but no confirmed multi-hop infections yet.