Asia-Pacific: WIC AI Report, China’s Agent Trust, Singapore’s Review Fatigue and Agent Safety, Korea’s Privacy Guide

On 3 November 2026, the World Internet Conference (WIC) published Frontier Artificial Intelligence Safety and Governance 2026: Risk Awareness and Collaborative Governance, prepared under its AI committee’s safety and governance programme to examine frontier AI risks, assess governance readiness and recommend coordinated responses.

The research report recommends extending safety evaluation from foundation models to deployed systems, distinguishing a model’s underlying harmful capabilities from the residual risks after safeguards are applied, and explaining the coverage and limitations of testing. For agents and open models, it proposes allocating responsibilities according to each party’s actual control over models, interfaces, permissions and use cases, supported by verification of safety commitments, incident sharing and coordinated response.

Businesses procuring or deploying agents can include models, tools, data access and operating environments in acceptance testing. For example, when a model moves from knowledge-based question answering to calling business systems, testing should be repeated under the new access conditions and execution permissions. Vendor safety statements should identify the relevant version, test conditions and known limitations; contracts should assign responsibility for changing permissions, suspending services and supplying investigation materials when anomalies occur. Safety commitments need to correspond to testable functions and clear delivery responsibilities so that businesses can assess the risks that remain as deployment expands.

The report’s governance readiness assessment covers 24 countries, using six dimensions and 22 secondary indicators, with an evidence cut-off of 31 August 2026; it explicitly states that its purpose is not to rank countries. The report series is intended as an annual public resource for monitoring governance practice. Businesses operating across borders should record the evidence cut-off and assessment methodology when using its regional analysis, then check the latest local rules before making market-access decisions.

On 4 November 2026, the Cyberspace Administration of China (CAC), together with relevant parties, proposed the Cooperation Initiative on the Innovative Development of the Intelligent Internet, bringing agent collaboration, cross-border data flows and emerging security risks into the international cooperation agenda.

The initiative proposes jointly developing open, non-discriminatory and transparent international standards for agent interoperability, promoting compatibility across interfaces, protocols, semantics and workflows, and establishing a trust framework spanning agent discovery, negotiation and execution. It also calls for measures against data poisoning and loss of control over agents, alongside risk information sharing and emergency cooperation. Data flows remain subject to safety, compliance and respect for differences between jurisdictions. This is a cooperation initiative; the detailed technical standards remain to be developed.

Businesses building workflows across agent platforms should check whether authorisation travels with a task as interfaces are connected: can downstream systems verify the initiator, delegation relationship, permitted actions and validity period, and stop subsequent calls when the original authorisation is revoked? Procurement testing can cover task handoffs, expired authorisations and requests exceeding permissions to check how different platforms enforce the same restriction. Cross-border transfers of customer records or personal information also require checks on recipients, actual data flows and the applicable transfer conditions.

China’s Global AI Governance Initiative, released on 18 October 2023, had already set out directions for international standards and governance cooperation. This week’s initiative makes that cooperation more specific to the intelligent internet and agents working across platforms, addressing interface compatibility, task collaboration and traceability of responsibility. Businesses can monitor how subsequent standards and pilots define identity, authorisation and traceability evidence, and incorporate those elements into interoperability testing and vendor assessments.

On 3 November 2026, Singapore’s Ministry of Manpower (MOM) confirmed in a written parliamentary answer that closely monitoring and validating AI outputs can create additional cognitive demands, including mental fatigue, and said it would continue to examine AI’s effects on work and employee well-being.

MOM does not currently plan to revise the tripartite workplace mental health advisory specifically for AI-related tasks. Employers can use existing guidance to identify and manage work stressors and refer to workplace fatigue management guidance where fatigue risks arise. The answer highlights an easily overlooked issue: reviewing AI outputs itself requires sustained attention and judgement, and human oversight needs working arrangements that support it.

Businesses using human review in customer service, content moderation, investment research or clinical support should include review capacity in deployment planning: how many outputs must each person handle, which errors are hardest to detect, and is there still time to consult source materials during peak demand? Risk-based task routing, rotation and escalation can be assessed alongside missed errors, repeated checking and fatigue feedback. The effectiveness of human review should be judged by the problems reviewers actually identify and correct; the time, training and decision-making authority required for these roles should also be included in project budgets.

Singapore introduced the relevant tripartite workplace mental well-being advisory on 17 November 2020, establishing a framework for identifying work stress and supporting employees. This week’s answer explicitly brings AI output review within that workplace issue, with MOM stating that it will assess the evidence to determine whether more specific guidance is needed. As AI use expands, businesses can assess how workloads shift from producing information to verification and correction, and adjust staffing accordingly.

On 6 November 2026, the Cyber Security Agency of Singapore (CSA), Government Technology Agency of Singapore (GovTech) and Infocomm Media Development Authority (IMDA) jointly issued safety advice for personal AI agents that connect to applications and accounts and perform tasks on users’ behalf.

Using Personal AI Agents Safely and Securely: What Consumers Should Know advises users to limit actual access through application settings and approve important actions individually, while checking persistent memory, data sharing and training uses. It notes that chat instructions are insufficient to constrain account access; when anomalies arise, users should stop the agent and disconnect affected accounts both within the agent and at the relevant service provider. The advice is primarily intended for individual consumers.

Businesses allowing employees to use personal agents for work should first establish which work accounts they connect to, which files they can read and whether task memories are retained over time. Pilots can begin with low-sensitivity, read-only tasks, with write and execution capabilities approved individually; payments, outgoing emails and changes to business records need defined human confirmation conditions. Incident exercises should test stopping tasks, disconnecting accounts and revoking credentials, with necessary records preserved before clean-up. Deletion of data already stored in the cloud should be checked separately.

The outcomes of CSA’s agentic AI tabletop exercise, published on 15 October 2026, had already addressed runtime monitoring, execution environment hardening and access control. This week’s joint advice applies those issues to personal accounts and everyday operations. Businesses can use it to improve employee training and reviews of personal-tool access, checking whether products support permission limits, one-time approvals and authorisation revocation, and use exercise results to inform expansion into further work scenarios.

On 4 November 2026, South Korea’s Personal Information Protection Commission (PIPC) announced a revised Guide to the Standards for Ensuring the Safety of Personal Information (개인정보의 안전성 확보조치 기준 안내서), publishing the full November 2026 edition on 5 November with further explanations of API management, authentication and AI security testing.

The guide explains the existing safeguards standards, emphasising server-side API authorisation, data minimisation in responses and limits on bulk requests, and recommends disabling unused interfaces and revoking keys or tokens that are no longer needed. For AI tools used in vulnerability assessment and penetration testing, it calls on users to consider differences arising from model choice, version and configuration, and whether source code, system configuration information or personal information is transmitted to external services.

Businesses processing personal information in Korea, or providing AI security testing services to relevant institutions, should include the testing tools themselves in data-processing reviews: define in advance which code and samples may be uploaded, the environments that may be accessed and the scope of testing, and check external services’ retention and subsequent-use conditions. Interface acceptance tests should cover unauthorised data retrieval, repeated bulk requests and invalid credentials. After changing an AI model or testing tool, reassess test coverage and whether differences in the results affect the system’s security assessment.

PIPC amended the relevant notification on 1 July 2026; this week’s guide adds implementation explanations. The supplementary provisions and Article 18 specify that institutions operating public systems within the prescribed scope must conduct at least one vulnerability assessment and one external penetration test for each covered system annually, with prompt remediation; these provisions take effect on 1 January 2027. Relevant institutions can use the preparation period to define system scope, testing procurement and remediation responsibilities, and account for potential service disruption and data damage during testing.