Abstract Zero-shot anomaly detection (ZSAD) aims to identify unseen anomalies without relying on anomalous samples, enabling unsupervised generalization to novel domains. However, most existing methods ignore mutual guidance between text and visual modalities during the prompt learning, thereby struggling to maintain consistent alignment with text anchors at local scales and across multiple representation layers. To address these challenges, we propose BiPrompt, a bidirectional visual-text prompts framework that establishes mutual guidance between textual and visual modalities. The proposed framework adaptively refines both visual features and textual representations through collaborative cross-modal interaction, allowing the model to focus on semantically relevant regions while suppressing noise and spurious responses. Furthermore, a hierarchical alignment mechanism enforces semantic consistency across multiple abstraction levels, strengthening fine-grained correspondence between language and vision. Comprehensive experiments on seven industrial benchmarks demonstrate that BiPrompt achieves superior performance in both image-level anomaly detection and pixel-level localization. The results confirm its effectiveness, robustness, and generalization capability in zero-shot settings, providing new insights into the integration of vision–language modeling and anomaly detection.