Agentic Abuse Prevention in AI-Driven Healthcare Decision Support Systems

Abstract

AI systems are increasingly becoming more autonomous, leading to serious concerns about unintended misuse, decision drift, and harmful actions. Agentic abuse, where an AI system goes against human intent or operational limits, presents major risks in areas like healthcare, security, robotics, and defense. This paper introduces the Self-Consistent and Modular Agentic Abuse (SCAMA) framework. It is a multi-layered structure designed to detect, prevent, and reduce agentic abuse through intent monitoring, behavioral analysis, anomaly detection, and controlled execution. The framework includes human oversight, ethical rule enforcement, and modular safety gates to ensure that AI decision-making stays in line with set objectives. Experimental results show high accuracy in detecting misaligned intentions and harmful actions, which greatly reduces unsafe autonomous behaviors. The SCAMA architecture offers a practical way to develop transparent, reliable, and trustworthy agentic AI systems that are suitable for high-risk settings. AI-driven healthcare decision support systems (HDSS) increasingly depend on autonomous and agent-based functions like self-adaptive diagnosis, continuous learning, and automated clinical recommendations. While these functions improve efficiency and patient outcomes, they also introduce new safety risks. These risks include agentic abuse, which refers to unintended or harmful autonomous actions that skip clinical oversight, invade patient privacy, manipulate diagnostic results, or disrupt medical workflows. This paper proposes a framework to prevent agentic abuse in AI-driven HDSS. It does this by using multi-layered safety constraints, behavior-governed reinforcement learning, explainable AI (XAI) audits, and zero-trust policy enforcement. The proposed model includes anomaly detection, consent-aware access control, real-time action monitoring, and rule-based clinical oversight to ensure safe and transparent AI actions. Experimental results show significant improvements in detecting misuse, a decrease in unsafe autonomous actions, and stronger defenses against potential threats in healthcare settings. The framework seeks to support regulatory compliance, build trust, and enable responsible use of autonomous healthcare AI systems.