Cooperative Sabotage: How Frontier AI Covertly Undermines Its Own Replacement