Agentic AI Security: Threat Analysis, Defense Strategies, Security Evaluation, and Emerging Challenges
Agentic AI Security: Threat Analysis, Defense Strategies, Security Evaluation, and Emerging Challenges
Zahra Baaji Kazemi1 Sama Goya2
1) Department of Computer Engineering, Ar.C., Islamic Azad University, Arak, Iran -
2) Department of Computer Engineering, Islamic Azad University, Tehran, Iran -
IOI Code
:
AAHD-ABKGCZ
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
Agentic AI systems powered by Large Language Models (LLMs) and equipped with capabilities such as planning, tool use, memory, and autonomy are emerging as powerful infrastructures for automating complex workflows. However, moving intelligent systems from passive response generation toward autonomous decision-making and action substantially expands their security exposure. This survey reviews major threats and vulnerabilities in agentic AI through a structured taxonomy covering prompt injection and jailbreaks, autonomous cyber-exploitation and tool abuse, multi-agent and protocol-level threats, interface and environment risks, and governance and autonomy concerns. It then examines defensive strategies including prompt-injection-resistant design, detection and isolation, policy enforcement, capability restriction, sandboxing, runtime monitoring, and organizational security frameworks. The paper also analyzes the evolution of evaluation from task-completion capability toward reliability, safety, and control, emphasizing security-oriented benchmarks that capture failures throughout an agent’s execution trajectory. Finally, open challenges are discussed, including long-horizon security, multi-agent coordination, adaptive attacks, physical-agent security, and human-agent interfaces. The overall analysis highlights the importance of secure-by-design development and continuous evaluation for trustworthy deployment of agentic AI systems.
Keywords : agentic ai ai security prompt injection cybersecurity multi agent systems security evaluation