Agentic AI: Autonomous Agents and Multi-Agent Security

Abstract

This paper presents a conceptual review of Agentic Artificial Intelligence (AI), autonomous agents, and multi-agent system security. It discusses key security threats, including prompt injection, excessive tool permissions, sensitive data leakage, memory poisoning, insecure inter-agent communication, and cascading failures. The paper proposes a layered security framework incorporating identity management, least-privilege access, policy-controlled tool usage, secure agent communication, memory protection, monitoring, bounded autonomy, and human approval for high-impact actions. It also outlines an evaluation methodology using adversarial test cases and measurable security and performance metrics. This work aims to support the development of safer, more reliable, and secure agentic AI systems. The proposed framework is conceptual and literature-based; no prototype implementation or experimental results are claimed.