Asia-Pacific: China Publishes an Embodied AI Ethics Report, Proposes AI Technology Export Catalogue Changes and Introduces Pre-export Registration; Hong Kong Updates Banking AI Supervisory Guidance; Singapore Publishes Agentic AI Security Exercise Findings; Japan Reports on the First Japan–Korea AI Safety Technical Exchange

On 13 October 2026, China’s National AI Standardization General Group published the “Research Report on Ethical Governance of Embodied Intelligence”, proposing risk classification and tiering, lifecycle governance and technical safeguards for the ethical risks of embodied AI.

The report is a governance research publication and does not, in itself, establish new mandatory certification or deployment approval obligations for enterprises. It identifies six categories of ethical risk, covering human autonomy, physical safety, privacy, discrimination, mistaken identity and accountability and traceability, and assigns five risk tiers, R1–R5, according to the severity of harm. Governance spans project initiation and design, development and integration, deployment and activation, operation and service delivery, and service termination and decommissioning. Technical safeguards include constraining device actions, controlling the collection of sensory data, and providing identity cues and exit functions.

Enterprises procuring or deploying robots for material handling, visitor guidance or eldercare can incorporate site safety testing and accountability records into procurement acceptance. Define permitted actions, operating areas and data collection boundaries for each task, then test whether the system can constrain movement, stop or transfer control to a human when people approach unexpectedly, instructions conflict, connectivity fails or the model makes an incorrect judgement. Review camera and microphone data collection, cloud transmission and subsequent training uses separately, to prevent suppliers from expanding data use on the grounds that it is necessary for device operation. Revalidate safety boundaries after changes to the model, control software or task scope. During operation, retain linked records of relevant versions, key instructions, action outcomes and human interventions, and allocate contractual responsibility for fault handling, update notifications and data deletion at retirement. These are deployment recommendations based on the technical safeguards and lifecycle governance described in the reporting; specific testing, recordkeeping and retention arrangements still depend on the use case and applicable rules.

The Ministry of Industry and Information Technology’s “Guiding Opinions on the Innovative Development of Humanoid Robots”, published on 2 November 2023, already called for safety risk assessment across complete robots, software and algorithms, alongside research into ethical standards. The new report further specifies risk categories and governance stages for embodied AI. The next verifiable development is work on national standards, including “Artificial Intelligence — Guidelines for Ethical Governance of Embodied Intelligence”, explicitly mentioned in the reporting. Relevant questions include how risk-tier criteria, validation methods and accountability and traceability will be set out in concrete texts; the status of those standards and their implementation arrangements remain to be verified. For regional comparison, product-related high-risk classification under the EU’s Artificial Intelligence Act (AI Act) requires examination of the AI’s role as a safety component, coverage under product legislation, and third-party conformity assessment conditions. This serves a different purpose from the report’s ethical risk tiers. Enterprises operating across regions can reuse test evidence, while recording ethical risk assessments and local legal classifications separately and checking when the relevant provisions apply.

On 16 October 2026, China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) issued Announcement No. 48 of 2026, seeking public comment on revisions to the “Catalogue of Technologies Prohibited or Restricted from Export from China”, proposing 23 additional AI-related technology entries and simultaneously introducing temporary pre-export registration for the technologies listed in the annex.

The proposed prohibited and restricted entries remain in draft form, while temporary pre-export registration is already in operation. Registration covers overseas transfers, licences for overseas use, and changes that place ownership or actual control with an overseas entity; exports already under way but not yet completed are also covered. Large models trained on China’s core data and subsequently used to provide services would be prohibited from export. Large models trained on China’s important data and subsequently used to provide services would be restricted, subject to exemptions in the accompanying notes. For a large model meeting any one of three criteria—compute of ≥10²⁶ floating-point operations, ≥2 trillion parameters, or a global top-10 position in the relevant public evaluations at the time of export—its weights, key checkpoints, source code and core technologies for training optimisation and inference acceleration would also be restricted. Separate proposed controls for agents use criteria covering planning, collaboration and tool use. The tool-use entry requires all of the following: support for integration with ≥200 third-party tools, tool-selection accuracy of ≥90%, and an average end-to-end success rate of ≥90% for coordinated tasks involving at least three types of tools.

Model developers, agent solution providers and overseas deployment teams should consider maintaining classification records covering technology assets, control criteria, recipients and delivery methods. Model assessments should record compute, parameter counts and the relevant rankings at the time of export, while also tracing training data sources. An important-data share of ≤0.1% alone is insufficient to establish an exclusion: the conditions in the notes concerning contribution to capabilities, safeguards and non-reproduction in outputs must also be checked. Agent projects require assessment of specific capabilities and combinations of criteria, rather than tool counts or base-model size alone. Link classification decisions to cross-border delivery approvals for weights, code, checkpoints and technical documentation. Complete registration before exporting technologies within the temporary registration scope, and review deliveries already started but not completed. Supplier contracts can allocate responsibility for classification materials, registration support and change notifications; reassess scope when versions, capabilities or delivery methods change. The registers and approval processes above are deployment recommendations, while registration itself is an operational requirement.

When the Catalogue was adjusted on 15 July 2025, MOFCOM already distinguished export prohibitions, licensing for restricted exports and contract registration for freely exportable technologies. This update adds temporary registration ahead of formal revision. The next milestones are the consultation deadline of 15 November 2026 and the final scope and implementation date; registration does not replace other applicable procedures. Open-source status and deployment methods also require separate assessment. The draft annex excludes AI technologies already in the public domain and provides for automatic licensing facilitation for export arrangements such as new open-source releases of restricted technologies and self-built, self-operated computing centres overseas. Regarding arrangements after the revised Catalogue formally takes effect, the official Q&A explains that overseas local deployment of restricted technologies falls within scope, while delivery through APIs, SaaS or similar service models falls outside controls on large-model technology exports. Article 53 of the EU AI Act, by comparison, exempts qualifying open-source models from only some documentation obligations, while copyright-policy requirements remain. Enterprises therefore need to assess China’s export-side arrangements separately from model governance requirements in the destination market; open-source status is not a uniform exemption across regions.

On 15 October 2026, the Hong Kong Monetary Authority (HKMA) issued updated “Supervisory Guidance for Responsible Adoption of Artificial Intelligence”, setting out revised AI governance and risk management expectations for all Authorized Institutions.

The published supervisory guidance specifies lifecycle controls proportionate to risk and materiality and calls on institutions to upgrade their policies, procedures and controls in a timely manner. Inventories of approved use cases should cover internally developed, third-party and embedded AI, with use outside the inventory restricted. End-to-end testing should take place before deployment and material changes. Material or higher-risk use cases should undergo independent validation and include a kill switch or equivalent mechanism to enable suspension, access revocation and prevention of further actions. Institutions should also maintain auditable records across the lifecycle, with retention arrangements aligned with relevant data protection, confidentiality and retention requirements.

For banks bringing credit assessment support tools or agents able to modify customer records into production, we recommend linking inventory registration to production access approval and identifying embedded AI features in procured software. For material or higher-risk uses, include independent validation reports and suspension drill results in deployment acceptance, test whether credentials can be revoked and subsequent actions blocked after suspension, and record who approved acceptance of residual risks. At procurement, agree conditions for authorising the use of customer data in training, change notifications and evidence export. Operational records should link model versions, authorisations, execution outcomes and human interventions to support investigation of anomalies. These are implementation recommendations based on the guidance and agentic AI governance concepts; specific configurations still need to reflect the risks of each use case.

The guidance supersedes the “High-level Principles on Artificial Intelligence” dated 1 November 2019. The “Concept Note on Governance of Agentic Artificial Intelligence”, published on the same day, remains a preliminary discussion document. Its exploration of an industry-level registry for material agents does not constitute a specific supervisory requirement. The HKMA has announced that a detailed discussion paper will follow, providing a basis for tracking subsequent requirements. Regional implementation arrangements also differ. The Monetary Authority of Singapore (MAS) issued its “Guidelines on Artificial Intelligence Risk Management” on 7 October 2026, covering all financial institutions and specifying phased implementation milestones of 7 October 2027 and 7 October 2028. Hong Kong’s updated guidance addresses Authorized Institutions and calls for timely control upgrades. Groups operating across regions can share inventories and validation evidence, while managing applicable entities and implementation plans separately.

On 15 October 2026, the Cyber Security Agency of Singapore (CSA) published the “Outcome Statement – Tabletop Exercise on Agentic AI Security and Control at the Singapore International Cyber Week 2026”, summarising the scenario-based exercise it co-organised with Google on 14 October and the discussion of system-level safeguards.

The publication reports tabletop exercise findings and policy discussions; it introduces no new mandatory obligations for enterprises. The statement emphasises system-level safeguards outside the model, including continuous runtime monitoring, hardened execution environments and strict access controls. Real-time, deterministic guardrails can block unauthorised high-impact actions, complemented by asynchronous monitoring of agent reasoning and tool use. For highly privileged tasks such as cybersecurity monitoring, participants also highlighted the risks of over-reliance on a single model and proposed independent verification and model diversity.

When deploying IT operations agents that can execute commands or access production environments, we recommend expressing permitted tools, target resources and action boundaries as executable policies, checking identity, authorisation and risk before invocation. Set blocking rules independent of prompts for actions such as deleting records, modifying configurations or exporting sensitive data, and define when human approval is required. Procurement acceptance can assess whether the platform supports execution environment isolation, credential revocation and export of tool-call evidence. Exercises should test response paths for agents bypassing restrictions, multiple agents making successive tool calls, and monitoring failures. These are deployment recommendations informed by the exercise and technical frameworks; specific access, approval and retention configurations still depend on the use case.

On 20 May 2026, Singapore had already published observations from its government–Google AI Agents Sandbox, identifying risks involving indirect prompt injection, privacy and human oversight. The latest scenario-based exercise brought policymakers, critical infrastructure leaders and safety and security experts together to examine system safeguards further. Follow-up areas include progress on the threat-intelligence sharing, pre-deployment testing and incident-response collaboration listed in the statement; the announcement provides no common implementation timetable. On the same day, the Hong Kong Monetary Authority (HKMA) published “Supervisory Guidance for Responsible Adoption of Artificial Intelligence”, specifying independent validation and suspension control expectations for material or higher-risk use cases. Both publications can inform control design, but their status differs: Singapore’s statement offers exercise recommendations, while Hong Kong’s guidance sets supervisory expectations for Authorized Institutions. Enterprises operating across regions still need to check applicable requirements separately.

On 16 October 2026, the Japan AI Safety Institute (J-AISI) reported on its first joint technical workshop with the Korea AI Safety Institute (Korea AISI), held on 8–9 October to exchange work on AI safety evaluation, sectoral applications and security technologies.

The event was a technical cooperation and research exchange; the announcement published no common standard or new enterprise obligation. Thirty researchers, engineers and other representatives participated, focusing on evaluation methods that account for multilingual and multicultural contexts. Japan presented its JAI-Trust project to build generative AI safety evaluation benchmarks, alongside work on healthcare, robotics, conformity assessment and data strategy. Korea shared model and system evaluation work and multimodal red teaming—adversarial testing to identify safety weaknesses across different types of input. The announcement published no joint test scores or certification mutual-recognition arrangements.

For enterprises deploying customer service or document recognition systems in Japan and Korea, we recommend separate Japanese- and Korean-language test sets, with reviewers familiar with local language and culture assessing bias, misleading information and refusal criteria. For multimodal inputs, also test whether combinations of images and text change safety performance. When procuring models, request results broken down by language and input type, record model and test-set versions, inference settings, scoring criteria and remediation retest results, and rerun tests after changes to models, prompts or image-processing pipelines. These preparations can support supplier comparisons and decisions on deployment scope. They are deployment recommendations informed by the workshop topics, rather than acceptance standards jointly prescribed by Japan and Korea.

On 29 May 2026, J-AISI had already published discussions of JAI-Trust, explaining its aim of building generative AI safety benchmarks and evaluation methods. The latest workshop brings domestic benchmark-development experience into bilateral technical exchange. Subsequent developments to check include whether the institutes publish reusable test sets, methodological documentation or further cooperation arrangements. Internationally, the evaluation best practices described by the UK AI Security Institute (UK AISI) emphasise aligning inference settings and scoring criteria when comparing results, alongside appropriate recordkeeping. This complements the Japanese and Korean focus on language and cultural adaptation: enterprises can share evaluation processes and record formats, while maintaining local test sets and deployment decisions for each market. Evidence reused across borders should specify the languages, scenarios and limitations to which it applies.