South Korea's state-run internet security agency is rewriting its rulebook for AI agents. The Korea Internet & Security Agency, ...
South Korea's state-run internet security agency is rewriting its rulebook for AI agents. The Korea Internet & Security Agency, which sits under the Ministry of Science and ICT, told Reuters on Tuesday that it is preparing a revised version of its AI Security Guide, first published last year. The updated guide will focus on security issues that arise as companies roll out agentic AI services, and it will include a checklist to help manage those risks.
The trigger is deployment rather than theory. Companies are putting AI agents into production that operate with limited human oversight, and the written guidance on securing them has not kept pace. A national agency revising a document is a small administrative act on its own. What it indicates is that agent autonomy has moved from a research question to a supervisory one.
The AI Security Guide is being revised, not written from scratch
KISA is not starting with a blank page. The AI Security Guide already exists and was first published last year. What the agency described to Reuters is a revised version of that document, reworked around a class of system that has become common in the months since.
That distinction matters more than it sounds. A new regime would need consultation, drafting and a political decision about scope. A revision ships faster and lands on organisations that already know the document. It also means the baseline is a year old, written before the current wave of agentic deployment, so the update is closing a gap rather than getting ahead of one.
It is worth defining agentic AI plainly here, because the term carries most of the weight in this story and is used loosely elsewhere. An agent is a system given an objective rather than an instruction, allowed to pick its own steps, and permitted to act on the outside world through tools it can call. The autonomy is not a claim about intelligence. It describes how much happens between a human decision and an effect, and every step in that gap is somewhere a control has to sit.
What a checklist actually commits anyone to
The concrete addition is a checklist. KISA said the revised guide will carry one to help organisations manage the risks that surface when agentic AI services go live. That is worth reading precisely rather than generously.
A checklist is a self-assessment instrument. An organisation works through it item by item and ends up with a record of having considered each point. It is not a licence, an audit or an approval gate, and nothing in the reporting suggests KISA intends it as one.
For background, security governance usually arrives in this order. The checklist comes first because it can be written quickly and applied voluntarily. Enforcement, if it follows, waits until the checklist has been in circulation long enough that regulators know which items mattered and which were noise.
Physical AI is the clause worth watching in these AI systems
The most consequential line in the story is also the most heavily hedged. KISA added that the guide could incorporate common control measures for physical AI, meaning systems able to interact with real-world devices and machinery. The agency said could, not will, and the difference should survive being repeated.
That clause carries weight because the failure modes differ. When a software agent goes wrong it writes something incorrect, calls an API it should not have called, or leaks data. When an agent with hands goes wrong, it moves something. Recovery is a different problem, and so is the question of who answers for it.
KISA also drew a line most coverage will blur. The agency said the guide addresses risks from agentic AI systems broadly, rather than targeting high-performance AI models such as those implicated in the Hugging Face breach. Regulators rarely go out of their way to decouple a policy from a recent incident, which suggests the work predates it.
For teams already running agents in production, the practical read is narrow but usable. A checklist written by a national security agency is a fair proxy for what a regulator will eventually ask, and those questions are answerable now: which actions an agent takes without a human, what it can reach, what is logged, and how it is stopped.
If your team is putting agents into production and wants a view on what to write down before a regulator asks, atxsoft.com can help you work through it.
Frequently Asked Questions
What is agentic AI?
An AI system given an objective rather than a fixed instruction, which chooses its own steps and can act through tools. KISA's concern is that these systems operate with limited human oversight.
What is KISA's AI Security Guide?
A guidance document first published last year by the Korea Internet & Security Agency, which operates under the Ministry of Science and ICT. KISA told Reuters it is preparing a revised version.
What will the updated guide cover?
Security issues that arise as companies roll out agentic AI services, and a checklist to help manage those risks.
What is physical AI?
Systems capable of interacting with real-world devices and machinery. KISA said the guide could also incorporate common control measures for them, though it stopped short of committing to it.
Is this a response to the Hugging Face breach?
No. KISA said the guide is intended to address risks from agentic AI systems more broadly, rather than being targeted at high-performance models such as those implicated in that breach.
When will the updated guide be published?
A publication date was not given in the reporting, and the contents of the checklist have not been released.
References
