Son Ki-wook outlines security requirements for AI agents

Son Ki-wook, president of the Korea Cybersecurity Academy and professor at Seoul National University of Science and Technology, has outlined critical requirements for the commercialization of AI agent security. Speaking at the PRIBIT AI Tech Summit, Son emphasized that security for AI agents must move beyond simple error prevention to addressing the risks of actual system manipulation and autonomous actions.

Key security focus areas include protecting 'memory' to prevent data contamination and managing 'tool calls' via the Model Context Protocol (MCP) to control how agents interact with external systems. Son argued that security products must provide transparency by maintaining evidence flows of an AI's decision-making process, allowing for human oversight and the ability to immediately halt actions when risks are detected.

Furthermore, Son noted that for AI agent security products to be commercially viable, they must balance protection with operational efficiency. Companies will increasingly demand that security measures account for latency, cost, and overall usability rather than just the number of attacks prevented.