To address the difficulty of existing backdoor attack methods for face recognition systems in achieving a balance among attack success rate, visual concealment, and cross-model transferability, this paper proposes a covert backdoor attack method for face recognition based on an Attention-Guided Hybrid Trigger. The proposed method combines a low-perceptibility elliptical visual trigger with a CBAM-guided semantic perturbation to construct a visual–semantic hybrid trigger mechanism. A multi-objective loss function is introduced to jointly optimize attack effectiveness, visual concealment, and cross-model transferability. Meanwhile, the mechanism of the hybrid trigger is analyzed from the perspective of feature space, revealing the reasons why the elliptical visual trigger enhances concealment and why the CBAM attention mechanism improves transferability. Experimental results on the LFW, CelebA, and CASIA-WebFace datasets with representative face recognition models, including FaceNet and ArcFace, demonstrate that the proposed method achieves a high attack success rate while maintaining the benign accuracy of the model with minimal impact. Compared with typical backdoor attack methods, the proposed approach achieves superior performance in terms of cross-model transferability and visual stealthiness. Ablation experiments further validate the effectiveness of the visual trigger, semantic trigger, and CBAM attention module in improving attack performance. Feature-space analysis reveals that the hybrid trigger can guide the stable migration of high-level discriminative features in the model, thereby enhancing the stability and generalization capability of the backdoor attack. Furthermore, this paper discusses the limitations of the proposed method in complex real-world scenarios and physical-domain applications, and provides prospects for future research on physical-domain backdoor attacks against face recognition systems and corresponding defense mechanisms. The findings of this study can provide valuable references for the analysis of backdoor attack mechanisms and the development of security defense technologies for face recognition systems.