NextAI+ Praxis--:----UTCOn August 20, 2026, China’s Measures for Cyber Data Security Risk Assessment formally took effect, bringing important data processors into an annual risk-assessment, reporting, and ongoing verification mechanism. On the same day, Korea’s National Assembly passed an amendment to the Personal Information Protection Act, creating a special pathway for using personal information in AI development where there is public or social necessity and the use is approved through regulatory review.
On August 20, 2026, the Cyberspace Administration of China (CAC), the Ministry of Industry and Information Technology (MIIT), and the Ministry of Public Security (MPS) formally brought the Measures for Cyber Data Security Risk Assessment into effect, establishing the cyber data security risk-assessment regime, annual assessment requirements for important data processors, and mechanisms for submitting risk-assessment reports. The Measures were promulgated by the three authorities on June 18, 2026 and took effect on August 20, 2026.
The Measures for Cyber Data Security Risk Assessment define cyber data security risk assessment as activities involving risk identification, risk analysis, and risk evaluation of the security of cyber data and cyber data processing activities, and establish a data-security management process covering assessment, reporting, inspection, and remediation. The Measures focus on three mechanisms. First, important data processors must conduct a risk assessment every year; where the security status of important data changes materially and may cause adverse impact, the affected portion must also be assessed in a timely manner. Second, assessments may be conducted internally or entrusted to a third-party assessment institution, but the commissioning relationship, allocation of responsibilities, and authenticity of the report must be clearly defined. Third, the Measures establish a reporting mechanism: after completing the annual assessment, important data processors must submit the risk-assessment report as required by the competent authority; where the competent authority is unclear, the report must be submitted to the provincial cyberspace authority or the national cyberspace authority.
For enterprises involved in AI training, model invocation, and intelligent-application operations, the first impact falls on data-asset identification and risk-assessment processes. In the past, data governance in AI projects often focused on lawful data sourcing, access permissions, and personal information protection. The Measures further require enterprises to consider how the data itself is classified and what risk status it has under China’s national data-security framework. Enterprises processing important data need to determine which datasets, business databases, model-training data, or agent operating logs fall within the scope of risk assessment, and maintain records that explain data flows, processing activities, and security measures.
In AI deployment scenarios, enterprises need to pay particular attention to continuous management once data enters the model lifecycle. For example, when enterprises in finance, energy, transportation, healthcare, and other sectors use industry data to train models or build knowledge bases, a one-off pre-deployment security review may not be sufficient to cover subsequent changes. Model retraining, dataset expansion, supplier replacement, cross-border invocation, and system-architecture changes can all alter the security status of important data and may therefore trigger the need to reassess whether another risk assessment is required. Enterprises should therefore connect risk assessment with model-version management, data catalogs, supplier management, and system-change processes, rather than treating it as a standalone data-compliance document.
At the same time, report-retention and external-verification requirements will affect how enterprises prepare for audits. The Measures for Cyber Data Security Risk Assessment require important data processors to prepare risk-assessment reports and retain them for at least three years; relevant authorities may inspect and verify the authenticity and accuracy of those reports. Enterprises using third-party cloud services, large-model APIs, or external data processors need to determine in advance whether suppliers can provide data-processing records, system logs, and information on security controls to support subsequent assessments and verification.
This regime was not introduced specifically for AI. It forms part of the broader evolution of China’s data-security regulatory system from principle-level requirements toward operational mechanisms. The Data Security Law of the People’s Republic of China established classified and graded data protection and risk-monitoring mechanisms; the Regulations on Network Data Security Management further detailed requirements for network data processing activities; and the Measures for Cyber Data Security Risk Assessment add procedures for risk identification, annual assessment, and regulatory reporting. The practical change after implementation is that important data security management moves beyond an internal-control requirement into a process of periodic assessment and regulatory inspection.
Compared with other regions, China’s approach places greater emphasis on continuous assessment of data-processing activities themselves and regulatory reporting. The EU AI Act primarily regulates specific AI systems through risk classification, technical documentation, and conformity assessment, while the United States relies more heavily on sector regulators to set requirements for particular fields. China’s data-security approach instead uses data classification, important data protection, and risk assessment as foundational mechanisms. For enterprises deploying AI systems across regions, this means that data-governance requirements can differ even when the model architecture is the same: the European market may focus more on high-risk system assessment and transparency requirements, while the Chinese market requires additional answers on data sourcing, data classification, risk assessment, and regulatory reporting pathways.
On August 20, 2026, Korea’s Personal Information Protection Commission (PIPC) announced that the National Assembly had passed an amendment to the Personal Information Protection Act establishing a special mechanism for the use of personal information in AI technology development. The amendment allows lawfully collected personal information to be used for AI development under specified conditions and following regulatory review. The amendment is currently an enacted but not yet effective legislative amendment; it must still complete the promulgation process and will take effect according to the statutory timetable.
The amendment changes the legal basis for using personal information in AI development rather than removing restrictions on personal information used for AI training. It creates a new special-use pathway for circumstances in which anonymized or pseudonymized data alone is insufficient for AI development: where the AI development has public or social necessity, the Personal Information Protection Commission has reviewed and approved the use, and enhanced safeguards are applied, previously lawfully collected personal information may be used for AI technology development.
The mechanism has three main components:
For AI enterprises, the primary impact is on training-data acquisition and model-development processes. Previously, enterprises conducting AI training in Korea relied more heavily on anonymization, pseudonymization, or renewed authorization. The amendment introduces a regulatory-approval pathway under which some data resources that were previously difficult to use may become available for AI development. But because the mechanism depends on PIPC review and cannot be invoked solely through an enterprise’s own judgment, it should not be understood as meaning that “existing personal data can now be used directly for AI training.”
At the data layer, enterprises need to revisit the sources and legal bases for training data. For example, if a medical AI company wants to use historical clinical data to improve a disease-prediction model and anonymization would remove features the model requires, the enterprise may in future need to assess whether the project meets the conditions for the special provision and prepare materials explaining data necessity, risk controls, and security measures. For multinational enterprises, this also means that Korean datasets may follow a different use pathway from data originating in other regions and need to be managed separately within global data-governance architectures.
At the model-development layer, enterprises need stronger traceability for training data. If an enterprise relies on a regulator-approved pathway to use personal information for model training, retaining only the final model parameters may be insufficient to explain the lawfulness of data use. The enterprise may also need to answer which data entered the training set, whether the data source was lawful, which projects received approval, and whether model training remained within the approved scope. This affects data-catalog management, training-pipeline records, and model-version management.
At the operations and audit layer, enterprises need to prepare in advance for regulatory engagement. Although the amendment has not yet produced a detailed application process, its design requires enterprises to demonstrate “why personal data is necessary,” “why anonymization is insufficient,” and “how risks to individual rights and interests will be reduced.” AI project data-impact assessments, supplier allocation of data responsibilities, and internal approval processes may therefore need to add review points specifically for training-data use.
This reform sits at the stage where Korea’s AI data governance is moving from principle-level discussion toward concrete mechanisms for data use. Previously, PIPC had already interpreted its guidance to allow, under certain conditions, the use of legitimate interests as a legal basis for processing publicly available data for AI model development, while requiring enterprises to assess purpose, necessity, and impact on individual rights and interests. The new amendment addresses a different question: how to establish a supervised data-use channel where public or anonymized data is insufficient to support AI development.
Compared with the EU and China, Korea is adopting a middle path of “permitted use + regulatory approval.” The EU General Data Protection Regulation (GDPR) relies more heavily on existing legal bases such as legitimate interests and consent to determine whether personal-data processing is lawful, together with restrictions in higher-risk contexts. China relies on the Personal Information Protection Law, data-security regimes, and generative AI management rules to establish data-processing boundaries. Korea’s amendment does not create a comprehensive AI data regime comparable to the EU framework; instead, it introduces a dedicated AI-development pathway within the existing personal information protection system.
From an enterprise deployment perspective, the core issue reflected in this change is not “whether data is open,” but whether the enterprise can explain why the data use is necessary and demonstrate that risks are controlled. Enterprises developing models, medical AI, speech AI, vision AI, or industry-specific large models in Korea will need to monitor the implementing rules for the approval mechanism and how it interacts with cross-border data transfers, model-supplier responsibilities, and global training-data governance. The amendment has passed the National Assembly, but detailed implementation rules have not yet been published. Enterprises can begin mapping data assets and preparing approval materials, but should not treat the amendment as an already effective rule that permits direct use of personal information for AI training.
Cite as · AI Governance Weekly · 26 August 2026
If you want both columns delivered together, four times a year, in one quiet email — leave an address. Otherwise just bookmark this page.