<p>Detecting malicious accounts on social media is becoming increasingly challenging as adversaries evolve from simple spammers to sophisticated actors capable of camouflaging their network topology. Existing unsupervised detection methods predominantly rely on explicit follower/followee structures. However, these topological signals are often manipulated by attackers (e.g., through link farming), compromising the reliability of structure-based Graph Neural Networks (GNNs) and obscuring the latent coordination among disconnected malicious users. To overcome the limitations of topology-dependent approaches, we propose BAGL (Behavior-Affinity Dual-Graph Learning), an unsupervised framework that unifies explicit structural interactions with implicit behavioral correlations. Unlike prevailing methods that treat user behavior merely as isolated features, we construct a novel Behavior-Affinity Graph that explicitly models the latent similarities among users from three complementary dimensions: statistical action distributions, interaction-driven influence patterns, and semantic content consistency. This graph unveils hidden connections between coordinated actors who share behavioral signatures despite lacking direct social links. Specifically, BAGL employs a dual-view graph encoding mechanism to extract representations from both the heterogeneous interaction graph (explicit view) and the behavior-affinity graph (implicit view). To ensure robustness against structural noise, we introduce a cross-view mutual information maximization objective that aligns structural and behavioral semantics, effectively disentangling genuine user intent from manipulated connections. Furthermore, a dual-graph reconstruction module and a soft clustering regularizer are jointly optimized to enforce local connectivity preservation and global semantic separability in a fully unsupervised manner. Extensive experiments on three real-world benchmark datasets, Cresci-2015, Twibot-20 and Twibot-22, demonstrate that BAGL outperforms representative unsupervised baselines, improving F1-score by up to 2.44% on the challenging Twibot-20 benchmark. On the large-scale and severely imbalanced Twibot-22 benchmark, BAGL still achieves the best F1-score among unsupervised methods and shows strong transferability of the learned representations. These results validate that integrating behavior-affinity priors is an effective component for identifying camouflaged malicious accounts in open-world social networks.</p>