South Korea Moves AI Safety From Models to Machines

South Korea is preparing an updated security guide for autonomous AI agents, marking a shift in AI governance from evaluating what models can generate to controlling what deployed systems are allowed to do.

The Korea Internet & Security Agency, which operates under the country’s Ministry of Science and ICT, told Reuters on September 15 that it is revising its AI Security Guide. The new version will focus on risks created when companies deploy agentic AI services that can use tools, access systems and act with limited human supervision. KISA said the guide will include a practical checklist and could establish common controls for “physical AI” systems that interact with machines and devices.

What changed

The move matters because autonomous agents introduce a different security problem from conventional chatbots. A model that produces a bad answer may cause confusion or reputational damage. An agent with credentials can modify a database, send a message, purchase an item or trigger an operational process. If connected to industrial equipment, vehicles or robots, the same permission failure could have physical consequences.

KISA has not announced a final rule or enforcement timetable. The agency’s language indicates guidance rather than a completed binding regulation, leaving open questions about which companies will be covered, how compliance will be assessed and whether controls will vary by sector or risk level.

Reporting by Seoul Economic Daily, published September 14, described the initiative as a broader effort to limit execution privileges and establish accountability across developers, service providers and users. That framing suggests the eventual guide may address the entire deployment chain rather than placing responsibility solely on model makers.

Why it matters

The practical center of gravity for AI safety is moving toward runtime control. Permissions, identity management, tamper-resistant logs, human approval and rapid shutdown mechanisms may become as important as benchmark scores or pre-release evaluations. This is particularly significant for enterprises adopting agents inside finance, customer service, software development and industrial operations.

South Korea is also positioning itself as an early test case for “physical AI” governance. The Ministry of Science and ICT has separately identified AI-agent security platforms and safeguards for autonomous vehicles, intelligent robots and drones as national priorities. That combination links software security policy to the country’s manufacturing and robotics ambitions.

The initiative could influence companies beyond South Korea if its checklist becomes a procurement requirement or a reference point for international standards. It may also expose a basic tension: stronger approval gates and narrower permissions reduce the damage from misbehavior, but they can make automation slower and less economically attractive.

What remains uncertain

The central question is whether the guide will produce measurable controls or remain voluntary best practice. It is also unclear how regulators will distinguish harmless workflow automation from systems capable of making high-impact decisions. Until KISA publishes the revised document, the announcement is best read as a policy signal: AI deployment is becoming an operational-security problem, not merely a model-quality problem.