New research challenges the assumption that AI safety primarily depends on model design, finding instead that operational guardrails and interaction rules in production environments fundamentally determine whether AI systems cooperate or exploit each other.