Analisis Kualitatif Kerentanan AI Agent terhadap Serangan Prompt Injection dan Strategi Mitigasinya

Abstract

The development of Large Language Model (LLM)-based AI agents in enterprise environments is driving increased digital automation while expanding modern cybersecurity risks. One of the most discussed threats in AI security developments is prompt injection, a natural language-based instruction manipulation technique that can influence AI agent behavior and trigger corporate data leaks. This study aims to analyze AI agents' vulnerability to prompt injection attacks, identify their impact on corporate data security, and examine security mitigation strategies based on global industry standards. The study used a descriptive qualitative method with a threat modeling approach and contemporary cyber case studies. The results show that AI agents are highly vulnerable to direct and indirect prompt injection, especially in systems connected to tool-calling, cloud storage, corporate email, and internal databases. Analysis based on the OWASP Top 10 for LLM Applications and MITRE ATLAS frameworks shows that prompt injection is closely related to sensitive information disclosure, excessive agency, and data extraction in enterprise operational environments. Mitigation strategies such as context isolation, input validation, role-based access control, adversarial testing, and strengthening AI governance are used to reduce the risk of exploitation of modern AI agents. The research implications point to the need to develop AI security standards that are more adaptive to natural language-based threats in enterprise digital ecosystems.