Communication technologies have significantly enhanced the functionality of DC microgrids, transforming them into complex cyber–physical systems. While distributed control schemes offer flexible and scalable coordination among multiple units, they also increase the system’s exposure to cyber threats. This article proposes a novel intelligent attack surface assessment framework for DC microgrids based on adversarial learning. A multi-agent reinforcement learning (MARL) algorithm is employed to emulate intelligent attackers, enabling systematic exploration of the attack surface for vulnerability assessment and the generation of sophisticated attacks that can bypass detection. Simultaneously, a data-driven attack detector is iteratively trained using operational data to identify these stealthy attacks. The detector provides feedback to the attacker, enabling the generation of new sophisticated attack patterns. Through this adversarial interaction, the proposed framework not only performs a dynamic assessment of the system’s attack surface but also advances the design of a more robust and adaptive detection scheme. As a result, a wider range of sophisticated attack vectors in converter-dominated systems can be identified. Experimental results validate the effectiveness of the proposed cybersecurity framework.