Metal surface defect detection serves as a crucial rule for ensuring the quality of industrial metal production. Nonetheless, various defects are prone to occur during metal processing, and even minor defects can negatively impact the performance of high-precision metal components. Despite the fact that the YOLO series detectors perform well in general object detection, there remain major challenges on the condition in that directly applying them to metal defect detection. This paper proposes LAC-YOLO, an efficient defect detection framework built on YOLOv8, for accurate and rapid identification of various defects. First, a lightweight multi-scale feature extraction module LEBottleneck is designed to project features into high-dimensional latent space through nonlinear transformations, enhancing feature representation while reducing model parameters. Second, a re-parameterized Aligned Feature Fusion Module (AFFM) is introduced to improve the representation of defect area details on the metal surface by aligning semantic expressions of adjacent layers while ensuring inference efficiency. Third, the classification-guided regression module (CGHead) enhances positioning accuracy by dynamically integrating category information into localization prediction while reducing computational complexity through deep separable convolutions. Experimental results illustrate that LAC-YOLO achieves 78.1% mean average precision (mAP) at 106 FPS on NEU-DET and 75.7% mAP at 110 FPS on DEFECT-DET, surpassing the baseline by 4.1% and 1.2%, respectively. These results validate the effectiveness and reliability of LAC-YOLO in industrial defect detection.