Security Framework for Autonomous Agentic AI Systems

This survey formalizes agent architectures as stateful 5-tuples to address security risks where natural language acts as both input and executable code. It establishes a 6-dimensional trustworthiness taxonomy to manage the expanded attack surface created by recursive reasoning and tool execution.