In this paper, we propose high-speed, accurate algorithms for detecting hazardous Web pages. Our algorithms automatically choose strings that appear especially in HTML elements of hazardous Web pages. We use these strings in combination as features of SVMs (support vector machines), and detect hazardous Web pages. Since our algorithms do not rely on the text parts of Web pages, they can detect Web pages that existing text-based algorithms have difficulty in detecting. By conducting a large-scale performance evaluation with real hazardous Web pages, we showed that the hybrid algorithms of our algorithms and existing text-based algorithms increase the precision of existing text-based algorithms alone by 9.3%.