Anthropic and EPFL show AI mind viruses can spread through persistent agent prompt files, while a one-paragraph warning cuts spread to near zero.