Glass, though ubiquitous, is difficult to recognize in an image due to its transparency. Fine-grained low-level features indicating the presence of glass, such as refraction and reflection, are weak and subtle. This causes difficulties for existing glass detection models in learning those features, pushing them to rely on more overt cues, especially the frame surrounding the glass. Consequently, they can be fooled easily by frame-like objects. Here, we propose a simple data augmentation scheme called Random Frame to address this problem. Random Frame inserts a frame into an image to create an area with a frame but no glass. The model will receive a penalty if it only relies on the frame. The performances of existing models on various datasets improve when Random Frame is applied while being trained. Our comprehensive experiments demonstrate that our data augmentation can make models utilize more low-level features with more confidence in their predictions.