The Korea Internet and Security Agency (KISA), South Korea’s principal internet security agency, said on Tuesday it is developing safety guidelines for agentic AI — that is, artificial intelligence programs that can operate without direct human supervision.
South Korea already has an “AI Security Guide,” a set of regulations formulated by KISA and the Ministry of Science in 2025 and officially implemented in January 2026. KISA officials said on Tuesday it was necessary to revise these guidelines to account for the growing popularity, and sophistication, of AI agents.
The South Korean agency was particularly concerned about “physical AI,” meaning drones and other real-world hardware controlled by AI systems.
WATCH — Innovation Council Action’s Taylor Budowich on AI Arm’s Race with China: Losing Is NOT an Option!:
Interestingly, KISA officials said they were aware of the “Hugging Face Incident” that prompted much of the recent anti-AI fervor in the United States but claimed their decision to update South Korea’s AI guidelines was not influenced by it.
“Hugging Face” involved a testing “sandbox” created by OpenAI in July. The idea was to establish a controlled, sealed environment in which hundreds of powerful AI models could be given various cybersecurity challenges and be evaluated.
According to OpenAI, the models “took actions that were misaligned with the goals of their assigned tasks,” as they “communicated through unauthorized channels, exploited vulnerabilities in shared infrastructure, gained internet access, and accessed third-party systems.”
The name of the incident came from a third-party platform called Hugging Face, which the “rogue” AI agents decided to hack into. The AI agents allegedly talked to each other on a message board they were not supposed to be using as they planned the Hugging Face hack, and they even had what amounted to an ethical debate about proceeding.
The AI agents ultimately decided that hacking the outside system was “unauthorized” behavior, but it would help them achieve the demonstration of cybersecurity prowess demanded by their programmers, so they proceeded with the attack.
WATCH — Substrate CEO James Proud on AI: We Must Stop China’s Domination of Tech Invented in United States:
KISA’s proposed revisions to South Korea’s security guide include restrictions on giving agentic AI access to certain software tools, requiring tamper-resistant logs of the decision-making process behind everything the AI agents do, requiring human approval for high-risk activity, and requiring service providers to include “real-time shutdown controls” for every agent, so they could be terminated instantly in the event of misbehavior.
South Korean companies accelerated their development of homegrown AI models after the Trump administration blocked the export of Anthropic’s state-of-the-art Claude Mythos model in June. South Korean tech firms and KISA had only gained access to Mythos a few days earlier.
The administration was concerned about the risks to national security and cybersecurity if Mythos, which is very adept at finding and exploiting software vulnerabilities, should fall into malevolent hands.
The ban was partly lifted a few weeks later, but South Korea took the Mythos ban as a warning that domestic companies should develop their own AI alternatives. One of the most urgent projects for a coalition of South Korean companies was Project Canopy, an attempt to develop enhanced cybersecurity protection with AI support.
“When we used to warn that the United States could suddenly cut off API access, people treated it as a hypothetical. Now that it’s actually happened, most countries are scrambling, but Korea moved early enough that we’re still able to create a response team,” a South Korean IT expert told Korea Joongang Daily (KJD) in June.
Read the full article here