South Korea AI conference says control key to agent safety

May 26 (Asia Today) -- As artificial intelligence evolves into autonomous agents capable of making decisions and taking action, experts in South Korea said Tuesday that the next phase of AI competition will depend less on performance and more on safety, control and trust.

The International Association for AI and Ethics hosted the 2026 AI Safety Compass Conference at Dreamplus Main Hall in Gangnam, Seoul, under the theme of AI agent security and corporate strategies for building trust.

Jeon Chang-bae, chairman of the association, said in opening remarks that humans and animals had long been the only beings with autonomy, but AI is now reaching a stage where it can act autonomously.

"As AI autonomy increases, the issues of control, safety and trust will become even more important," Jeon said.

Related

• South Korea launches public-defense AI alliance to pursue 'AI G3' goal
• South Korean AI power diagnostics system heads to Germany
• Japan moves to reshape moral education for AI era

Kim Myung-joo, head of the AI Safety Institute, said core principles for managing agent AI risks include granting minimum authority, ensuring traceable identities and securing auditability.

Kim said AI agents should receive only the permissions needed for specific tasks and should not be allowed to connect to unverified external services or install unapproved plug-ins.

He also stressed the need for a "kill switch" that can immediately block abnormal AI behavior.

"When an agent performs abnormal actions, it must be possible to disconnect it from outside systems and reduce its authority," Kim said. "Humans must be directly involved at every important decision point."

Lee Jae-hyung, head of the AI security response team at the Korea Internet & Security Agency, said AI is shifting from being only a target of cybersecurity to becoming an active participant in security operations.

"In the AI era, organizations must redesign their structures and decide how much work they should delegate to AI," Lee said.

He said major risks include AI misjudgment, uncontrollable decision-making and dual-use applications.

Lee also warned that AI could be used to automate hacking and cyberattacks, including by generating smishing messages that exploit human psychology and by lowering language barriers for attackers.

Speakers cited Anthropic's Claude Mythos Preview model as an example of how advanced AI can become both a hacking tool and a defensive instrument. Lee said preliminary results disclosed Friday showed the model had identified about 10,000 vulnerabilities among partner organizations.

"We have entered an era in which powerful AI can be both a means of hacking and a tool for defense," Lee said. "Security must be treated as essential to using safe and trustworthy AI."

-- Reported by Asia Today; translated by UPI

© Asia Today. Unauthorized reproduction or redistribution prohibited.

Original Korean report: https://www.asiatoday.co.kr/kn/view.php?key=20260526010007520