Leveraging LLM to Strengthen ML-Based Cross-Site Scripting Detection