LG AI Research's AI Ethics Seminar: The Rise of Agentic AI and Security Threats

As LLM-based generative AI evolves beyond a simple conversational partner into Agentic AI that sets its own goals and uses tools, the security paradigm is shifting from content safety to behavioral safety and integrity of purpose.

According to the Agency–Security Paradox, the higher an AI's autonomy, the greater the system control privileges granted to it, which increases the risk of leading to actual security incidents. This means AI can cause tangible harm—such as account changes or file deletion—beyond simply producing wrong answers.

Threats to Agentic AI occur at each stage of the cyclical loop of perception, memory, reasoning, and action.

• Perception: Identity spoofing and data injection by malicious agents (IPI)
• Memory: Memory Poisoning targeting RAG/Vector DBs
• Reasoning: Intent destruction, goal drift, and bypassing human oversight (HITL)
• Action: Tool Misuse and privilege violations, code execution (RCE) risks