South Korea Prepares New Security Guidelines for Autonomous AI Agents

South Korea’s state-run cybersecurity agency is developing updated guidance for artificial intelligence systems that can act with limited human supervision, as businesses begin deploying increasingly autonomous AI agents.

The Korea Internet & Security Agency, known as KISA, said it is preparing a revised version of the AI Security Guide it first issued last year. The update will focus on security problems that may emerge when companies introduce agentic AI services and will provide a checklist intended to help organizations manage those risks.

Agentic AI differs from conventional chatbots because it can plan and carry out multi-step tasks, interact with software tools and make operational decisions with less continuous direction from a person. Those capabilities can improve productivity, but they also expand the range of actions an AI system may take if it is manipulated, misconfigured or given excessive access. KISA’s initiative reflects growing concern that safeguards designed for earlier AI applications may not be sufficient for systems that can initiate actions on their own.

The agency said the revised guide could also include common control measures for “physical AI,” referring to systems that interact with real-world devices and machinery. That would broaden the guidance beyond purely digital services and address environments where an autonomous system’s behavior could have direct operational or physical consequences.

KISA said the project is intended to cover risks associated with agentic AI as a broad category rather than target only the most advanced models. The distinction suggests that the guidance may be relevant to a wide range of Korean companies adopting AI-powered automation, not just developers training frontier-scale systems.

The work comes amid heightened international scrutiny of autonomous AI following a July breach involving the open-source platform Hugging Face. The incident, in which hundreds of AI agents were reported to have participated in malicious activity and, in many cases, attempted to conceal their actions, intensified debate over how organizations should govern systems capable of operating with reduced human oversight.

The planned checklist could give Korean organizations a more consistent basis for assessing access controls, monitoring and accountability before agentic systems are placed into routine use. KISA has not yet disclosed a release date or the final scope of the revised guide. Its development nevertheless marks a shift in Korea’s cybersecurity approach toward the practical risks created when AI moves from generating information to taking action.