AI 안전 다이제스트

2 August 2026 – 10 September 2026 · 143 source cards
10 September 2026
  • National Sports Promotion Corporation earned NIA's 'AI Ethics Compliance Institution' certification mark, with over 70% of staff completing AI ethics training and new internal guidelines on public interest, safety, and accountability. [1]
  • SPRi published a study on growth paths for top-tier AI talent, defining core competencies for researchers versus engineers based on interviews and surveys, and recommending expanded K-12 entry pathways and stronger university research infrastructure. [2]
  • SPRi proposed a strategy for cultivating 'entrepreneurial AI talent', based on a survey of 715 advanced AI professionals and startup founder interviews, recommending graduate-level entrepreneurship training and startup-friendly ecosystem support. [3]
9 September 2026
  • The Financial Services Commission/FSS held a CISO meeting with the entire financial sector on frontier AI-driven cyber threats, urging enterprise-wide IT risk management and rapid patching of security gaps ahead of network-separation deregulation. [1]
  • The National Assembly Research Service issued a report calling for a precautionary regulatory framework for child and teen chatbot risks, citing emotional over-dependence and high-risk behavior tied to generative AI chatbots. [2]
  • The National AI Strategy Committee said it will strengthen its role as an execution-focused control tower, including improving the evaluation system for the Independent AI Foundation Model (Dokpamo) project and building a K-AI cybersecurity framework. [3]
8 September 2026
  • MSIT launched a policy study on AI glasses, aiming to curb test-cheating and unauthorized recording risks and develop usage guidelines, while the Personal Information Protection Commission separately begins work on data-handling principles for AI glasses. [1]
  • Korea's amended Personal Information Protection Act creates a special exception allowing use of personal data beyond its original collection purpose for AI development when anonymization/pseudonymization is infeasible, subject to safeguards and review by the Personal Information Protection Commission; it takes effect March 9, 2027. [2]
  • The National AI Strategy Committee and MSIT agreed to revise evaluation methods and criteria for the Independent AI Foundation Model (Dokpamo) program, with a cybersecurity-specialized model prioritized short-term and frontier-capable independent models pursued as a mid-to-long-term goal. [3]
7 September 2026
  • MSIT and IITP are budgeting about 3 billion won to set up a joint AI research lab in Europe next year, tasked partly with ethics/safety verification research and helping Korean firms respond to EU AI Act requirements, linking it to the existing New York AI Frontier Lab. [1]
  • At a defense AX and security conference, experts called for integrated governance across the Ministry of National Defense, MSIT, and the National Intelligence Service to build AI-based defense systems beyond K-RMF and zero-trust as drones, autonomous weapons, and cyber warfare converge. [2]
4 September 2026
  • MSIT selected a Naver Cloud-led consortium as the security-specialized track of the Independent AI Foundation Model project (Dokpamo), with 33 industry-academia-public institutions building two 700B-class MoE models for cyber offense/defense. [1]
  • Korea's AI policy command structure was reshuffled with new presidential and national AI committee appointments, and experts are pushing to redesign the Dokpamo project and link it to a frontier-model initiative. [2]
  • The same reshuffle also puts under review whether AI Framework Act regulations will take effect after next year's grace period ends, alongside complementary measures. [2]
3 September 2026
  • MSIT selected the Naver Cloud consortium to build a cybersecurity-specialized AI foundation model, backed by government-supplied GPUs with a mid-term evaluation after 5 months deciding continued support. [1][2][3][4][5]
  • The Financial Services Commission finalized a second emergency easing of network-separation rules, letting 75 financial firms test AI and security SaaS tools to detect vulnerabilities, with results to feed into future AI security guidelines. [6]
  • The Prime Minister's Office held its second field meeting on AI administrative innovation, where officials proposed expanding immunity for AI use and revising incentive/responsibility structures for civil servants. [7]
2 September 2026
  • Korea's Fair Trade Commission issued a revised notice requiring prior substantiation for AI and new-tech product performance claims, effective September 3, and cut the evidence-submission extension window from 30 days to 15 days, with ad-suspension orders for non-compliance. [1]
  • The National AI Strategy Committee met with a European high-level labor-policy delegation to discuss a socially legitimate national AI strategy, tying the government's 'Everyone's AI' push to employment, inequality, and public-trust concerns. [2]
1 September 2026
  • KISA and MSIT will deploy an AI penetration-testing agent in this year's second-half cyber crisis response drill, having it analyze company websites to craft intrusion scenarios against consenting firms from October 19 to 30. [1]
  • Lawmaker Ahn Sang-hoon and 10 others proposed amending the AI Basic Act to require AI operators to clearly notify users when they are interacting with a conversational AI system, with violations subject to fact-finding and corrective orders; the bill is under legislative notice through September 12. [2]
  • A Korea Capital Market Institute analysis found generative AI use in the financial sector expanding beyond internal support into investment advice, credit assessment, and customer consultation following relaxed network-separation rules, urging regulators to track performance and feed real-world risk experience into future policy. [3]
August 2026 (23 days)
31 August 2026
  • KISA and KCA held an 'AI Privacy Challenge' where university students used generative AI to find vulnerabilities in personal-data processing systems while institutional security staff practiced real-time threat detection and incident response. [1]
  • Ha Jung-woo took office as standing vice-chair of Korea's National AI Strategy Committee, pledging to coordinate AI governance across government and industry and stressing the need for sovereign AI systems in defense and public sectors. [2]
30 August 2026
  • Korea's presidential office restructured its AI policy leadership, naming Lee Hae-min as AI future planning secretary, reappointing Ha Jung-woo as full-time vice chair of the National AI Strategy Committee, and confirming Bae Kyung-hoon as science minister overseeing the Independent AI Foundation Model project and national AI computing center. [1][2][3][4]
29 August 2026
  • MSIT selected SK Telecom, KT, and Kakao as operators for the 'Everyone's AI' free public AI service, with each firm developing GPU-supported search, booking, and payment features to launch later this year. [1]
28 August 2026
  • MSIT's second-round evaluation for the Independent AI Foundation Model project (Dokpamo) advanced Upstage, SK Telecom, and LG AI Research while eliminating benchmark leader Motif Technologies, triggering a formal appeal and demands to disclose detailed expert-panel scores amid transparency concerns. [1][2][3][4]
  • MSIT selected the SK Telecom, Kakao, and KT consortium to build 'Everyone's AI,' the free national public chatbot, after evaluating six competing bidders, with GPU support and a beta launch planned for later this year. [5][6][7][8]
  • MSIT and NIA narrowed a cybersecurity-focused AI foundation model program to two consortia, SK Telecom-Upstage and Naver Cloud-LG CNS, who will validate models on real security data and jointly assess performance and safety with academic and public partners. [9]
27 August 2026
  • Korea's AI service user-protection public-private council held its second meeting to review illegal/harmful content risks and regulatory gaps, with Kakao presenting its Kanana Safeguard guardrail model for safety verification. [1][2]
  • The International AI Ethics Association urged Korea's National Education Commission to fully reconsider using AI for first-pass grading of essay-type exam answers, citing bias, lack of explainability, hallucination risk, and inadequate remedy procedures. [3]
  • MSIT and NIA released 29 AI training datasets (about 35.44 million items, ~1.56 trillion tokens) built under the Independent AI Foundation Model project by five selected teams, after quality and safety verification, for free public use via AI Hub. [4][5]
26 August 2026
  • Korea's Independent AI Foundation Model project advanced three teams — SK Telecom, LG AI Research, and Upstage — past its second evaluation round based on benchmark, expert, and user scoring, while Motif Technologies was eliminated. [1]
  • ETRI joined the Agentic AI Foundation under the Linux Foundation to help shape agentic AI standards such as MCP, working alongside roughly 250 member companies on accuracy, security, and governance workstreams. [2]
  • Korea's National Assembly Research Service flagged implementation gaps in the newly effective AI and Data-Based Administration Act, citing unresolved issues in data quality, AI reliability/accountability, and governance/security ahead of its August 28 enforcement date. [3]
25 August 2026
  • AhnLab warned that "memory poisoning" attacks can corrupt an AI agent's long-term memory to distort future decisions, and Korea's privacy watchdog said it will draft AI agent guidelines and related legislation in the second half of 2026. [1][2]
  • The government finalized Korea's AI ethics principles, built on three values and seven principles (human-centeredness, privacy, fairness, accountability, safety, reliability, transparency), explicitly naming users as responsible actors alongside developers, though the code remains a non-binding voluntary norm under the AI Basic Act. [3][4][5]
  • ETRI joined the Linux Foundation's Agentic AI Foundation (AAIF), a roughly 250-member group including Google and OpenAI, to help set international standards for agentic AI protocols such as MCP. [6]
24 August 2026
  • MSIT finalized Korea's new AI ethics principles, built on three values and seven principles, explicitly designating users, government, and civil society as "AI actors" alongside suppliers, and requiring human control over AI agents plus stronger accountability for large AI providers. [1][2][3][4][5][6][7]
  • Commentary flagged that because the revised principles remain a non-binding self-regulatory code, expanded civic oversight and eventual legislation are needed to make them effective against harms like job loss and privacy violations. [7][1][2][3][4][5]
22 August 2026
  • Korea amended its Personal Information Protection Act to allow use of personal data for AI development when pseudonymized or anonymized data is insufficient, subject to review by the Personal Information Protection Commission and mandatory prior risk assessment for sensitive data; the change takes effect 6 months after promulgation. [1]
21 August 2026
  • South Korea's National Assembly passed an amendment to the Personal Information Protection Act allowing lawful use of personal data for AI development when pseudonymized or anonymized data is insufficient, subject to enhanced safeguards and Personal Information Protection Commission review; use of sensitive or unique identifier data requires prior risk assessment and disclosure, with the law taking effect 6 months after promulgation. [1]
20 August 2026
  • South Korea's National Assembly passed an amendment to the Personal Information Protection Act allowing original personal data to be used for AI training when public interest is recognized, subject to review by the Personal Information Protection Commission, mandatory prior risk assessments for high-risk data processing, and public disclosure requirements. [1]
19 August 2026
  • MSIT's second-round evaluation of the Independent AI Foundation Model project (Dokpamo) passed Upstage, SK Telecom, and LG AI Research to the next stage while eliminating Motif Technologies, based on benchmark, expert, and user-evaluation scores. [1][2][3][4][5]
  • Critics noted the government did not disclose team-by-team scores or rankings from the tax-funded Dokpamo evaluation, despite using a 40/35/25-point benchmark-expert-user rubric. [6]
  • MSIT is reviewing a broader restructuring of the Dokpamo and related AI programs, including future competition rules, resource allocation, and whether a fourth selection stage will proceed. [1]
18 August 2026
  • South Korea's Independent AI Foundation Model (Dokpamo) project eliminated Motif Technologies in its second-stage evaluation despite top benchmark scores, advancing Upstage, SK Telecom, and LG AI Research to the third round; the government said a review with global evaluator Artificial Analysis found no evidence of benchmark-targeted overfitting. [1][2][3]
  • The Korea Institute of Finance called for financial firms to establish AI governance frameworks covering fairness, transparency, consumer protection, and data security, urging regulators to set minimum explainability standards for black-box models while letting firms implement governance autonomously via sandboxes. [4]
17 August 2026
  • MSIT's revised second draft of AI ethics principles reportedly removed or weakened anti-discrimination provisions covering gender, age, and disability, as well as developer/provider accountability language present in the first draft. The Korean Women's Associations United criticized the draft for shifting bias/discrimination responsibility onto users and demanded explicit non-discrimination and inclusion principles with clear duties for each actor across the AI lifecycle. [1]
14 August 2026
  • MSIT and KISDI unveiled a draft revision of Korea's AI ethics principles, updating the 2020 standards into 7 principles (including safety, reliability, human dignity, privacy, and fairness) built on 3 core values, designed as non-binding soft-law guidance rather than new regulation. [1][2]
13 August 2026
  • South Korea's government proposed an amendment to the Personal Information Protection Act that would expand permissible use of personal data for AI development when enhanced safeguards are in place, subject to review by the Personal Information Protection Commission; the bill is currently under National Assembly deliberation after passing preliminary readings in May and July 2026. [1]
12 August 2026
  • The Board of Audit and Inspection found that public institutions have built duplicate and poor-quality AI training datasets despite over 1.6 trillion won in government investment, and recommended MSIT introduce prior data-sharing checks and unified quality standards. [1]
11 August 2026
  • KISA and IITP convened an industry council to shape information-security R&D priorities, focusing on AI agent security, zero-trust, and physical-AI risks. [1]
  • The Korea Institute of Finance issued a brief urging regulators to prepare consumer-protection rules for AI shopping agents before they reach retail financial customers, citing disclosure and conflict-of-interest risks. [2]
9 August 2026
  • MSIT announced an 'Agentic AI Initiative' covering physical AI safety, including plans for safety guidelines and sandbox support for humanoid robots after a forum raised concerns about malfunction risks from physical manipulation. [1]
7 August 2026
  • Korea lacks pre-release safety verification for AI models: MSIT's AI Safety Institute evaluated 42 domestic and foreign AI models but found no cases of pre-launch review, unlike the US system where firms submit models to government for pre-release assessment. [1]
6 August 2026
  • MSIT and KISA published an AI security red-teaming guide to help companies conduct pre-emptive risk checks on training data, core models, applications, and autonomous agent behavior. [1]
3 August 2026
  • Korea's Personal Information Protection Commission urged Samsung Electronics and Meta to build in privacy safeguards for upcoming AI glasses, such as recording-indicator alerts to protect bystanders, and plans to establish related privacy principles this year. [1]
2 August 2026
  • AI-Ethics-KR criticized Korea's autonomous vehicle law and the proposed physical AI special act for including vague "unavoidable" personal-data-use exemptions that prioritize industry competitiveness over clear standards or civil-society input. [1]