Federated Learning (FL) systems have been widely adopted across distributed clients to collaboratively train a shared model without exchanging raw data. However, FL systems are vulnerable to model poisoning attacks, where malicious clients send poisoned updates to deteriorate the performance of the global model. Most existing attacks either ignore the impact of benign client updates, or rely on multiple consecutive communication rounds, causing their attack impacts to diminish rapidly once the attack stops. This paper proposes the Durability-Optimized Model Poisoning Attack (DOMPA), a framework designed to maximize the durability of attack impacts. To implement DOMPA, we design a poison generation strategy, enabling the malicious clients to generate poisoned updates to induce a significant degradation on the global model. Furthermore, we develop a durability optimization mechanism that strategically aligns poisoned updates with benign updates to enhance the durability of attack impacts. Extensive experiments demonstrate that DOMPA is powerful and effective even against robust aggregation rules. Particularly, DOMPA achieves a higher reduction of accuracy while extending the durability of attack impacts over baselines. In addition, experiments in practical FL systems also disclose their vulnerability. Our code is available at https://github.com/Jiuh-star/DOMPA .