Deployment rules causally shape multi-agent AI safety as much as model choice does.
This paper introduces institutional red-teaming, a method to test how deployment rules (not just AI models) affect multi-agent safety.