Executive Summary

Korea is drawing up a national plan for AI safety, and companies have been given seats in the room where it is written. On September 30 the Ministry of Science and ICT said it had formed a public-private working group to prepare the plan, under the working Korean title '(가칭) 국가 AI 안전 생태계 조성 마스터플랜', and that the group would begin full operation in October. The ministry has published no English title; the National AI Safety Master Plan is the rendering used here. About 25 people take part: the ministry and related public agencies, Korean and global AI companies and their trade associations, and experts in technology, law and institutional design. The ministry wrote that it would settle the plan by December, after the group's discussions and consultation with other ministries.

What has been published so far is the frame. Three strategies carry it: building a national AI safety management system, strengthening the response capacity for advanced AI, and making AI safe in everyday life. What a company must hold, and what it must show to prove it, has not been written down yet. The date has already moved once. The ministry opened a public comment session on February 6 and said then that the substance would be announced in the first half of the year; the September 30 announcement wrote the deadline again as December. At a parliamentary audit on October 6 Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon repeated the December plan, and as of November 17 no meeting result or draft from the group has been made public.

Sections 1 through 5 take their facts and figures from Korea Policy Briefing's account of the ministry's press release, coverage of the parliamentary audit, two government publications issued in July, reporting on a bill tabled on November 2, and the text of the EU AI Act. On November 17 the Korea Policy Briefing press-release list for November 10–17 was checked again, and the one related release issued in the meantime was added. The attachment to the ministry's press release is posted only as a Korean word-processor file, which could not be opened directly; the figures inside it were confirmed against the Policy Briefing text and several consistent news reports. Section 6, which groups four kinds of records, and all of Section 7 are this article's reading.

Key Numbers

The first number is the size of the room where the plan is written. The middle two show how much money and how many people stand behind it today, and the last one is a level of demand the EU has already set.

Sources: Korea Policy Briefing, Digital Daily, EU AI Act Article 26.

About 25

People in the public-private working group

Government, public agencies, Korean and global AI firms, associations and experts in one room

224×

Gap between the AI performance and infrastructure budget and the AI safety and trust budget

₩6.149 trillion against ₩27.4 billion for next year, from Rep. Lee Ju-hee's audit material

38

Staff at the AI Safety Institute

Up from 26. At the same hearing the UK institute was put at more than 100

6 months

Minimum log retention the EU asks of companies deploying high-risk AI

Not a Korean requirement. It is a reference line already written down elsewhere

1

What the Ministry Set Up on September 30 Is a Working Group

The press release the Ministry of Science and ICT issued on September 30 is headed '민·관이 함께 국가 인공지능 안전 종합계획 수립 추진', which says that the public and private sectors will work together to draw up a national AI safety plan. What it reports is the formation of a body. The ministry set up a public-private working group to prepare the plan that carries the working title '(가칭) 국가 AI 안전 생태계 조성 마스터플랜', and said the group would begin full operation in October. The plan itself did not appear. The people who will write it did.

The schedule in the announcement runs to two lines. The group operates from October, and the ministry draws the master plan together by December, combining the group's discussions with consultation among other ministries. Three strategies were given as the frame: a national AI safety management system, a stronger response capacity against advanced AI, and AI safety in daily life. The group's remit is to review those three strategies and their core tasks, and to find new tasks or fill out existing ones in light of global AI safety policy and of industry and technology trends.

Lee Jin-su, who heads AI policy planning at the ministry, said the country had built a base for balancing AI innovation and safety, including the entry into force of the AI Framework Act, but that it had reached the point where national capacity against advanced AI risks needs strengthening. He added that the ministry would draw up a workable way of building the national AI safety ecosystem, reflecting what the field will accept and what global AI safety policy and technology are doing. The policy archive run by the Ministry of Foreign Affairs carries the same announcement, summarized as the establishment of a national AI safety policy framework that strengthens the ground of safety and trust alongside AI innovation.

Two names make this easy to misread. The '국가 인공지능 안전 종합계획' of the press release headline and the '(가칭) 국가 AI 안전 생태계 조성 마스터플랜' of its body point at one document. The first is the phrase used to report the news; the second is the provisional title the government has written down for the plan. When the document is settled in December, even the name may change.

2

Who Sits in the Room, and What Companies Do There

The group of about 25 breaks into four blocks. From government come the Ministry of Science and ICT and the National AI Strategy Committee. The public agencies are the AI Safety Institute, the National Information Society Agency, the Korea Information Society Development Institute and the Telecommunications Technology Association. Industry brings not only major Korean AI companies but global AI companies, small and medium firms and startups, and trade bodies such as the Korea AI and Software Industry Association and the Korea Startup Forum. The last block is made up of experts in AI technology and in law and institutional design.

The four blocks are not there to do the same job. The diagram below carries over the division of labor as the announcement set it out.

Public-private working group, about 25 people Government Ministry of Science and ICT National AI Strategy Committee Public agencies AI Safety Institute · NIA KISDI · TTA Industry Korean and global AI firms SMEs · startups · associations Experts AI technology Law · institutions What industry carries in Experience and views from the places where AI is built and used. What experts and agencies carry in A review of existing tasks and a search for new ones, on technical, policy and legal grounds. Field experience reaches the plan through this room. Companies outside it will be reading the December document.

▲ Pebblous original diagram. The participating bodies and the division of labor are carried over from the ministry's September 30 announcement.

The role the announcement gives companies is plain. It is a seat for carrying in what has been run into on the ground where AI is built and used. Reviewing tasks and finding new ones, by contrast, is written down as the work of the experts and the public agencies. What companies say in that room, then, can shift the yardstick the December document uses.

No list of participants has been published. Which companies took a seat, and what the phrase global AI companies points at, are not in the announcement either. How often the group has met, and what it discussed, is not known yet. Most companies that run AI in industrial operations, in manufacturing or energy, are likely to be outside the room, and for them the announcement is less a piece of news about taking part than a trailer.

3

No Law Fixes the December Date

December made the headline in report after report. The nature of that date is worth pinning down. The sentence in the September 30 announcement is that the ministry will settle the plan by December, after the group's discussions and consultation among ministries. No law nails the deadline down, and nobody carries a penalty for missing it. It is a schedule the government has written for itself.

This schedule has moved once already. On February 6, at the Bankers Association building in central Seoul, the ministry held a public comment session on the '(가칭) 국가 AI 안전 생태계 조성 마스터플랜' together with the Telecommunications Technology Association and the AI Safety Institute. The same plan, under the same name. Kim Myeong-ju, head of the AI Safety Institute, gave the opening presentation, and participants from the civic group People's Solidarity for Participatory Democracy and from academia asked for redress procedures for those harmed and for a sector-by-sector approach. The ministry said at the time that it would reflect the views raised there and announce the substance during the first half of the year.

Had the plan arrived in the first half, the September 30 announcement would have had no reason to write a deadline again. The diagram below strings together the public announcements and nothing else.

February 6 Public comment session Goal: first half of 2026 September 30 Working group announced Goal: plan by December October 6 Parliamentary audit December restated November 17 Results and draft No release found Only the public announcements are strung together here. What happened between the two stated goals was never published, and whether December holds or moves again cannot be read from what is public.

▲ Pebblous original diagram. All four points come from published announcements and news coverage.

The plan came up again on October 6, at the audit held by the National Assembly's Science, ICT, Broadcasting and Communications Committee. Bae Kyung-hoon, Deputy Prime Minister and Minister of Science and ICT, answered that the safety ecosystem master plan would be settled and published by December. At the same hearing Rep. Lee Ju-hee pointed to the distance between two numbers in next year's budget: ₩6.149 trillion for AI performance and infrastructure against ₩27.4 billion for AI safety and trust. She also noted that the AI Safety Institute had grown from 26 staff to 38, while the United Kingdom runs more than 100 and Japan aims at 200. Bae acknowledged that evaluations of high-impact AI and of AI safety are often commissioned to the British side, and answered that building up the institute's headcount and budget to secure that capacity is what matters.

One announcement did arrive in the meantime. On November 12 the Ministry of Science and ICT and the AI Safety Institute launched the "AI Safety Guardians," a panel of 80 citizens who will report cases of factual distortion, bias, security weaknesses and technical defects in generative AI through August 2027 (Korea Policy Briefing press release). It sits close to the third strategy, AI safety in everyday life, but the release does not mention the master plan.

That is as far as the record goes. December is a goal written in September and spoken again in October, and as of November 17 there is no public account of when or how often the group has met, or of what a draft contains. From outside, there is no ground for saying the date will hold and none for saying it will slip. Which is why reading it as settled, and hanging an internal schedule on it, is the risky move.

4

The Act Is in Force, but Where Is the Document That Says How?

This working group was not raised on empty ground. The Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation of Trust, Korea's AI Framework Act, took effect with its enforcement decree on January 22. The text now in force is the partial amendment promulgated on January 20, which has applied since July 21. The AI Safety Institute and the National AI Strategy Committee, both seated in the group, are bodies that live inside that framework. Because the law had already made the seats, what the September announcement describes is not a new body but the work of filling in the plan that goes into them.

The law that writes a duty and the document that writes how the duty is met do move separately, though. Two administrative notices from the ministry are the case in point. One sets out what an operator handling high-impact AI has to do ('고영향 인공지능과 관련한 사업자의 책무 고시'); the other sets out how the safety-assurance duty is to be met ('인공지능 안전성 확보 의무의 이행방법 등에 관한 고시'). Neither has an official English title, so the descriptions here are this article's own. As of October 14 neither notice had been issued (Maeil Ilbo, October 29). No announcement that they have been issued since is confirmed as of November 17. The two notices sit apart from the master plan due in December: they write how the duties of a law already in force are to be carried out.

The side that pointed at this gap was business. On October 29 the Federation of Korean Industries gathered its members' views and submitted 19 proposed improvements to the AI Framework Act to the ministry. The first asked for a line to be drawn on whether AI used only inside a company carries the same duties as a service sold outside it. A request to settle the notices carrying the compliance standards for safety-management duties quickly went in alongside it (Pebblous covered that submission separately: Korea's FKI Asks If AI That Never Leaves the Company Is Regulated, October 30, 2026). A company without a seat in the group can still send the same message through an association or a business federation.

More useful on the ground is what the same framework produced in July. On July 8 the Ministry of Science and ICT and the Korea Internet & Security Agency published two documents, a manual for responding to AI security threats and a guide to AI security red teaming. The manual sorts threats into data, model, agent, supply chain and high-performance model, and writes separate scenarios for eight industries: finance, healthcare, public administration, education, manufacturing and energy, telecommunications, law, and IT. It gives executives the threat types and the cases, and gives security and IT operators the criteria for diagnosis and response.

The red teaming guide that came with it writes down how to put together and run a team that hunts for weaknesses from an attacker's side. It walks through the whole course, from planning and composition to preparation, execution and the reporting of results, and attaches checklists, inspection tools and job descriptions for red team members. It states that it was written on the basis of ISO/IEC 42119-7, the international draft standard for AI red teaming. Both documents can be downloaded from the websites of the ministry and the agency.

Two volumes are already in hand while the December plan is awaited. Neither puts a duty on a company. That is part of why they are useful. They carry the government's own yardstick for what should be inspected, and records kept against that yardstick can be used again when a standard does arrive.

5

Factory-Floor Rules Reached the Assembly First

The clue to what a company running AI on an industrial site will be asked for came out somewhere else first. On November 2 Rep. Choi Soo-jin, with ten other members, tabled a partial amendment to the Act on Promoting Industrial Digital Transformation and the Use of Artificial Intelligence. It adds a new Article 12-2 requiring the Minister of Trade and Industry to draw up guidelines for developing and using industrial AI so that it is safe and trustworthy. The bill sits with the Trade, Industry, Energy, SMEs and Startups Committee.

The items news coverage reports for those guidelines read like operating procedure on a plant floor. Cybersecurity, access-rights management, data integrity and anomaly detection across the whole course from development through operation and updates are in. Where risk passes a set level, a means of human intervention or of stopping the system has to be in place, and there are provisions against core process information leaking or being used improperly. When an incident occurs, there has to be a system for a person to step in at once, record management that lets the cause be traced back, and a recovery procedure. Grounds for supporting small and mid-sized companies short of AI staff and security capacity, through diagnosis, consulting, applied security technology and pilot deployments, are written in as well.

Two caveats go on the record. First, the bill is at the tabling stage. Neither its passage nor its date of effect is settled, and clauses can change in committee. Second, this bill belongs to the Trade, Industry, Energy, SMEs and Startups Committee, while the master plan of Sections 1 through 3 belongs to the Ministry of Science and ICT. The two tracks move separately. There is no ground for expecting the items in an industrial AI guideline to land unchanged in the National AI Safety Master Plan.

5.1Requirements Already Written Down Abroad

While the Korean standard is empty, demands of the same character have already been written into text elsewhere. The EU AI Act tells a company deploying high-risk AI to assign human oversight to a person who has the necessary competence, training and authority. The same article has it keep the logs the system generates automatically, so far as those logs are under its control, and nails the period down at "at least six months".

Carrying that article over does not mean Korea will ask for the same level. It is a reference line, showing the shape a demand takes once it exists. Name a person for the role, put the authority to intervene inside the system, and hold the records that pile up automatically for a set period. All three overlap with the items in the tabled industrial AI bill.

6

Four Records to Keep Before the Rules Arrive

Starting to make records after the standard lands is starting late. A regime that weighs safety mostly asks not whether a thing is safe now but how it has been handled all along. The test result of a version that has passed, or the judgment made at the moment of a change, does not get restored afterwards. Within the overlap of the items seen in the previous sections, the four below are records that can be kept before any standard is settled, put into the language of data operations. From here on this is the article's own compilation, not a list the government has asked for.

6.1An Owner per Task, and the Line Where a Person Steps In

List the tasks AI has a hand in, and put a named owner against each one. What comes next matters more. Write down, in numbers and in names, which value crossing which line brings a person into the judgment, and who holds the authority to stop. Not a title on an organization chart but the person actually sitting in that seat. If a person has intervened, when they stopped it and what they were looking at stays in the record too. The human intervention the tabled guideline speaks of and the oversight assignment the EU article speaks of both take this record as given.

6.2Tests and Approvals, Logged per Version

Whatever goes onto the floor, model or rule, needs a version number on it. What was tested with that version and how it came out, and who approved putting it into service, stay as one bundle. Which set of data, from which point in time, was used for training and evaluation goes in with them. The red teaming guide issued in July asks for the course from planning to the reporting of results to be written down for the same reason. The question that comes later is rarely whether a test was run; it is what was tested, and on which data.

6.3What Was Rechecked After a Change

AI in production is not put up once and left. A model gets swapped, a threshold gets adjusted, a sensor is replaced, a connected system updates. That is also why the tabled guideline covers not only development but the whole course of operation and updates. Record, change by change, what was altered, what was checked again afterwards, and which items were passed over without a check. When the change history and the revalidation results sit apart from each other, the path back to a cause breaks the moment an incident happens.

6.4What the Customer Checked and Took

AI on an industrial site mostly sits between a supplier and a buying company, which makes the boundary of responsibility easy to blur. Write down at the point of handover which version went across and when, what conditions and limits were passed along with it, and who on the customer's side checked what before accepting. For the supplier it becomes the record that proves its own share. For the buying company it becomes the ground for tracing which premises the system it operates is standing on.

Keep the records in four separate places and they have to be matched up again at the moment of use. Bound together, they take this shape. Which version went onto which judgment in which process, when it was approved and by whom, what was tested on which data before that approval, and who stops the line when a value crosses it, all run through a single record. The version that went to the customer is confirmed in the same place. Begin matching those links only after an incident, and proof takes time even with the records in hand.

One property is common to all four. Every one of them can be written only at the moment the event happens. Who stopped the line, which data a test ran on, what was changed and what was looked at again, what the customer checked and took away: none of it comes back six months later. Whether the standard arrives in December or later than that, what can be handed over then is only what was built up before.

7

Why Pebblous Is Watching This Plan

A regime that asks for safety ends up asking for records. A regime that asks for records asks next where those records came from and how they changed. Lineage, versions and revalidation after a change are problems data quality work has handled for a long time, and that is where the four kinds of records set out in the previous section point. News like this is easy to file under policy; from the side that handles data, this is the part that stays.

Pebblous does not take part in this working group. It holds no designation, and there is no basis for saying any product of ours has been certified against any standard. What is worth writing down is that the questions that come up again and again in the work of preparing the data AI uses on industrial sites overlap with the items at issue here. A model with no record of which version of data trained it, a dataset nobody looked at again after it changed, an automated judgment with no owner written against it: all three are common.

So when the December document is read, three things are worth reading with it.

  • What it says has to be proved. A demand about a current state and a demand about a history take very different amounts of time to prepare for.
  • Who checks that proof. Self-inspection and outside verification call for records in different forms.
  • How far it applies. Whether the same items attach to the side that builds a model and to the side that buys one in is a fork in the road on an industrial site.

When the plan appears in December, some of the expectations written here will miss. The order, though, does not change. Records have to pile up ahead of the standard, and a company with none starts counting from the day the standard arrives.

R

References

Official Documents

Korean News Coverage

Earlier from Pebblous