Anthropic and EPFL show AI mind viruses can spread through persistent agent prompt files, while a one-paragraph warning cuts spread to near zero.
Because 'trust me' isn't a permission model for your AI coding agent.