Large language models are becoming both powerful security tools and high-value targets, creating new attack surfaces where multimodal exploits, jailbreaks, prompt leakage, and AI-assisted cybercrime are evolving faster than current defenses.