In the last two decades, the abundant amount of unstructured data available on the Internet is seen as vulnerable to cyber threats. It significantly impacts on individuals, businesses, governments, etc. Considering this, we present a new study for cyber threat detection via deep learning-based approach. To this end, we devise a model called CASKET to detect cyber threat from the input textual data as vulnerable or non-vulnerable. Our proposed model is novel in the sense that it captures both semantic and contextual knowledge from the input textual data to detect cyber threat. The empirical evaluation of this study is conducted on two benchmark datasets. Our proposed CASKET model performs better on both datasets and shows better F-score and Accuracy values as compared to the existing works and baseline methods.