Tag Multi-Agent AI

Researchers from Anthropic and EPFL have demonstrated self-propagating payloads that jump from one AI agent to another through editable system prompt files. The findings expose a new class of security threat unique to multi-agent AI architectures.