What do language models know and when do they know it: Explainable and trustworthy AI for global threat detection