Fault diagnosis is essential to ensure bearing safety in industrial applications. Many existing diagnostic methods require large scales of data from a full range of working conditions. However, the structure and working conditions differences between machines lead to significant variation in data distribution, making it difficult to diagnostic with unseen samples. To handle this, a novel physics-embedded digital twin-assisted bearings fault diagnostic framework (PE-DaT) under unseen working conditions is proposed, effectively leveraging the inclusivity of the denoising diffusion probabilistic model (DDPM, i.e. diffusion model) and the idea of digital twin method for unseen sample acquisition. In this digital twin-assisted diagnostic framework, a physics-embedded diffusion net (DiffPhysiNet) is proposed for working information embedding and fault mechanism integration. Specifically, signals under limited working conditions with extended dynamic information are taken as the input for forward noising process. Then, DiffPhysiNet reconstructs signals under extended working conditions by a reverse denoising process. A physics-embedded UNet (Physi-UNet) is designed to extract working information and fault mechanism during the reverse process. Ultimately, extensive experiments on two bearing datasets (BJTU-RAO and PU) validate the effectiveness of our method compared with the state-of-the-art baselines and the ablution test confirms the significant role of DiffPhysiNet.