Every AI Policy Ever · open in the interactive register
Draft / proposedProposed · Guidelines / framework
Guideline on Securing AI Safety (draft, Article 32 of the AI Basic Act)
인공지능 안전성 확보 가이드라인
Record KR-05
- Issued by
- Ministry of Science and ICT, distributed through the National Information Society Agency (NIA) and the Korea AI Software Industry Association (KOSA) help desk
- Jurisdiction
- South Korea
- Level · region
- National · East Asia
- Published / adopted
- 2026-01-22
- Last updated
- 2026-02
- Language
- Korean
Summary
Draft guidance for operators subject to the safety duty of Article 32 of Korea's AI Basic Act, which under Article 24 of the Enforcement Decree covers AI systems meeting all three criteria: training compute of 10^26 floating-point operations or more; the most advanced (state-of-the-art) AI technology in current use; and a risk that may broadly and seriously affect life, physical safety and fundamental rights. It explains how to determine the systems in scope and who bears the duty, risk identification, assessment and mitigation across the life cycle, a risk-management system for safety incidents, and reporting, including submission of the safety measures to the Minister within three months.
Language
No English version of this instrument was found. The English title and the summary are the register's own, written from the sources below; for the exact wording use the Korean text.
Safety topics
general-purpose / frontier AIincident reportingevaluation & testingcatastrophic risk
Open question
issuing_body is unsupported and the checker gave no evidence-backed fix; date_last_updated is unsupported and the checker gave no evidence-backed fix; the checker found something possibly newer: {'what': 'KOSA help desk version of Feb 2026 (최첨단 인공지능 안전성 확보 가이드라인)', 'where': 'sw.or.kr', 'note': 'not Issue #184
The record shows its last checked values until a maintainer decides; the decision is recorded in its history.
Status notes
The NIA version (file dated 22 January 2026, 71 pages) opens with a notice that this draft guideline is still being written and may differ from the final version to be published. It explains, and reproduces in full, a draft MSIT notification on the methods of implementing the AI safety obligation (인공지능 안전성 확보 의무의 이행방법 등에 관한 고시(안)), whose number is left blank. The draft notification requires a first report to the Minister within 24 hours of recognising a safety incident, an initial-action report within 7 days and a result report within 15 days of the first report, and new submissions within one month of a substantial change that increases risk or of recognising a new risk. Cumulative compute is to be calculated from training records and system logs, including pre-processing and fine-tuning, by theoretical, empirical-statistical or hardware-based methods. The guideline is a reference document that neither replaces the law nor imposes new duties, though supervisors may use it when checking compliance.
Local copy
Archived. The secondary source could not be downloaded (The site's robots.txt does not allow automated downloads of this page). Last attempt 2026-10-18 07:41:24 UTC
Official source (third-party copy, not published) archive/KR-05/2026-10-12/official.pdf
Retrieved 2026-10-12 04:43:43 UTC
SHA-256 c49c4fa002e60d21a4964fbea992e1ae2f3554063db6e5da262b1d74aeb32dd5
Verification
BAnother official source, such as an announcement by the issuer, confirms the title, date and status. checked 2026-10-18 · alternative-official-url-fetched
Refresh of 18 Oct 2026 under the rule that level A needs the instrument's own text or an official register. Stored and read: NIA's copy of the draft guideline (nia.or.kr, PDF, 71 pages, stored 12 Oct 2026 and unchanged when the archiver checked it on 18 Oct 2026; read through pdftotext, as the PDF's own text layer is garbled) and NIA's notice of 22 January 2026 distributing the AI Basic Act guidelines, stored in this run. The cover image shows the title, '2026. 1' and the logos of MSIT and ETRI, but the cover has no text layer, so the month cannot be quoted, and no page of the text gives a date for the guideline; the publication date therefore rests on NIA's notice (its date and the file name 260122), an announcement that counts as alt-official, and the record is level B. The PDF confirms the draft marking, the Article 32 and Decree Article 24 basis, the three criteria, the draft notification with its blank number, the compute-calculation methods, the chapter structure, the three-month submission rule and the 24-hour, 7-day and 15-day incident reports. The running header on every page misprints the title as '인공지능 안정성 확보 가이드라인'; the notice's file name gives the title correctly. MSIT as author rests on the cover logos and on the guideline's purpose of explaining MSIT's notification; no text names it as the author. The KOSA help desk's later version (draft notification No. 2026-126 of February 2026, title '최첨단 인공지능 안전성 확보 가이드라인'), on which date_last_updated '2026-02' and the KOSA part of issuing_body rest, could not be stored: the site's robots.txt does not allow automated downloads, and sw.or.kr is listed as not official. The archiver flagged NIA's notice for review because it barely mentions the guideline's title; it is the notice that distributes the guideline files.
Second check: Independent check by Claude Code (claude-sonnet-5-5), 2026-10-18: 3 field(s) supported, fixes applied to title, original_title.
History
- 2026-10-18verification.level: A → B — level computed from the evidence by tools/pipeline.py2026-10-18-refresh-level-rule-2 · Claude Code (claude-opus-5-5)
- 2026-10-14notes: Both the NIA version (file dated 22 January 2026, 71 pages) and the later KOSA v → The NIA version (file dated 22 January 2026, 71 pages) opens with a notice that — The incident-reporting timelines are in the stored NIA version (draft notification Article 8(3), page 12), not only in the KOSA version. Removed the statements resting only on the unstored KOSA version and the grace-period sentence (now in verification.note). Added the compute-calculation rule, the one-month submissions and the guideline's reference character (pages 5, 9 and 12).2026-10-14-refresh-translation-fixes · Claude Code (claude-opus-5-5)
- 2026-10-14summary: Draft guidance for operators subject to the safety duty of Article 32 of Korea's → Draft guidance for operators subject to the safety duty of Article 32 of Korea's — 'Frontier-level technology' became 'the most advanced (state-of-the-art) AI technology in current use', following Decree Article 24(1)2 (page 8) and the guideline's own gloss '최첨단(State-of-the-Art, SOTA)' (page 19). Added that all three criteria must be met and the duty-bearer question (pages 14 and 22).2026-10-14-refresh-translation-fixes · Claude Code (claude-opus-5-5)
- 2026-10-14original_title: 인공지능 안전성 확보 가이드라인(안) → 인공지능 안전성 확보 가이드라인 — '인공지능 안전성 확보 가이드라인(안)' was not verbatim: the title on the cover and in NIA's file name has no '(안)'; the draft marking comes from the page-2 notice and stays in the English title.2026-10-14-refresh-translation-fixes · Claude Code (claude-opus-5-5)
- 2026-10-14english_version: – → {"kind": "none found"} — No English version in the stored copies, and none found in a targeted search on 14 Oct 2026.manual-2026-10-14 · Claude Code (claude-opus-5-5)
Citation
Ministry of Science and ICT, distributed through the National Information Society Agency (NIA) and the Korea AI Software Industry Association (KOSA) help desk. Guideline on Securing AI Safety (draft, Article 32 of the AI Basic Act). 2026-01-22. https://www.nia.or.kr/common/board/Download.do?bcIdx=28987&cbIdx=99835&fileNo=15. Status: Draft / proposed; verified 2026-10-18 (level B). Local copy archive/KR-05/2026-10-12/official.pdf retrieved 2026-10-12 04:43:43 UTC, SHA-256 c49c4fa002e60d21a4964fbea992e1ae2f3554063db6e5da262b1d74aeb32dd5. Every AI Policy Ever, version 1.18 (snapshot 2026-10-18), record KR-05.
Improve this record
Know of an amendment, a status change or an error? Report news or an error on GitHub; the form opens with this record filled in.