Synthetic Voices, Real Threats: Evaluating Large Text-to-Speech Models in Generating Harmful Audio