A Safety and Security Framework for Real-World Agentic Systems