Anthropic, EPFL show AI 'mind viruses' hop agents via persistent prompt files
An Anthropic-EPFL preprint demonstrates self-propagating payloads that survive context wipes by writing into persistent prompt files (SOUL.md/MEMORY.md), infecting the next agent up to 55% of the time. A one-paragraph system-prompt warning cut spread to near zero, framing the risk as real but currently limited.