<p>Cone-beam computed tomography (CBCT) is widely used in radiotherapy (RT) for patient setup. However, full-view projection data are not always acquired for every treatment fraction, whereas two orthogonal projections are routinely available. In such scenarios, two-view CBCT (2V-CBCT) becomes a desirable solution for volumetric image reconstruction, which is critical for dose verification and accumulation. Existing supervised learning approaches rely on training with paired two-view projections and ground-truth volumetric images, which are often difficult to obtain in clinical settings. To overcome the limitation of acquiring paired training data, we propose a zero-shot 2V-CBCT reconstruction method based on a plug-and-play (PnP) diffusion anatomical prior. The prior is learned from volumetric images collected at the same anatomical site, without requiring paired projection data. Our method leverages a generative diffusion model and adapts the diffusion-prior-based inverse problem solver by replacing the standard mean estimate with a measurement-informed expectation. To improve computational scalability and avoid the high cost of backpropagation inherent in posterior sampling, we introduce a novel and efficient PnP reconstruction strategy from a maximum-a-posteriori (MAP) perspective. Specifically, we solve intermediate MAP subproblems using a single-step preconditioned gradient descent with an appropriately chosen step size. To reduce reconstruction artifacts arising from inconsistencies between simulated and real projection data, we further propose a boundary masking technique to alleviate the mismatch effects of implemented 3D cone-beam forward operator. Preliminary evaluations on pelvic dataset demonstrate the superiority of the proposed method over supervised 3D U-Net baseline. Our approach achieves higher PSNR and MS-SSIM, confirming its effectiveness. We introduce a novel zero-shot 2V-CBCT reconstruction framework that utilizes a diffusion-based plug-and-play anatomical prior. By eliminating the need for paired training data, the proposed method improves flexibility and generalizability, particularly in settings where accessible-view configuration varies across reconstruction tasks.</p>