When AI Acts: Hacking the Age of Agentic Systems

AI security has been a recurring theme at HITCON. As language models evolve from linguistic comprehension to agentic action, the security challenges of AI have grown substantially.

LLMs are inherently susceptible to adversarial attacks; vulnerabilities such as jailbreaking, prompt injection, and model alignment represent challenges that are theoretically difficult to eliminate. As Agentic AI bridges diverse systems, it not only inherits vulnerabilities from the underlying infrastructure but propagates them across all integrated environments, forming a complex and interdependent supply chain that significantly expands the attack surface.

Beyond this, the foundation model, training procedures, training datasets, and even Tool Use all emerge as new attack vectors, yet security continues to be deferred to the final stages of the development lifecycle, repeating a well-worn mistake from IT history.

How to embed Security-by-Design into AI system development in this new era has become the next critical challenge for the information security community.