Asia-Pacific: MAS Finalises AI Risk Guidelines and Clarifies Agent Authorisation; Australia Investigates HeyCyan; China Incorporates AI Risk Controls into Industrial Deployment

On 7 October 2026, the Monetary Authority of Singapore (MAS) issued the final Guidelines on Artificial Intelligence Risk Management for Financial Institutions, setting out its supervisory expectations for AI risk management and phased implementation dates.

The Guidelines cover all financial institutions and all forms of AI. They are finalised supervisory guidelines that have not yet taken effect, with implementation proportionate to scale and risk. They take effect on 7 October 2027, when institutions are to implement Sections 3–4 on board and senior management oversight, AI identification and inventories, and risk materiality assessment; Sections 5–6 on life cycle controls, capabilities and capacity are to be implemented by 7 October 2028. Institutions whose AI use is unlikely to have a material adverse impact on themselves, their customers or other relevant stakeholders if it performs poorly or becomes unavailable may be subject only to basic governance policies and procedures, to be implemented in the first phase. MAS also makes clear that appropriate controls for higher-risk use cases should be implemented as soon as practicable, rather than waiting until the end of the two-year transition period.

For implementation, institutions should translate preparation into deployment deliverables with clear acceptance criteria. First, link procurement records to business use-case inventories, identify AI embedded in material vendor services, and document the data involved, accountable owners, applicable jurisdictions and tools accessible to AI agents. Next, incorporate vendor integration into production approval, test against the institution’s own use cases and representative data, assess customer data exposure, erroneous outputs and unauthorised actions, and agree change-notification arrangements. For uses assessed as higher risk, such as lending and underwriting, arrange validation independent of the development and deployment teams. For agents capable of taking action, configure least privilege, human intervention and deactivation mechanisms within the system. During operation, set monitoring thresholds according to risk and retain approval, testing, access and incident-response records to establish an evidence trail for audit review. These deliverables span procurement, data flows, model validation and production operations, requiring coordinated delivery by business, risk and technology teams.

The process began with the consultation paper of 13 November 2025 and has now reached the final Guidelines and implementation-preparation stage. Consultation feedback led to two substantive changes: the full implementation period was extended from the proposed 12 months to 24 months, and the proposed dedicated cross-functional AI committee was replaced by flexibility to use existing structures capable of providing coordinated oversight. An official next step is already identified: MAS intends to consult further in 2027 on supplementary guidance needed for AI agents in financial services, although the specific rules have not been determined. For regional comparison, Annex III of the EU Artificial Intelligence Act classifies creditworthiness assessment of natural persons, excluding financial fraud detection, and risk assessment and pricing for life and health insurance as high-risk uses. MAS covers all AI use by financial institutions and calibrates controls according to risk. Cross-regional institutions should therefore share inventories and evidence infrastructure while recording EU statutory classifications and MAS risk assessment outcomes separately.

On 9 October 2026, Chia Der Jiun, Managing Director of the Monetary Authority of Singapore (MAS), used a public speech to outline runtime authorisation, pre-execution checks and accountability mechanisms for financial AI agents, and to report progress on the related industry framework.

The Safeguards for Agentic Finance at Runtime (SAFR) framework described in the speech proposes governance checks before an action is executed: first, establish agent identity and authorisation; then check the proposed action against the user’s mandate, institutional policies and risk limits; and retain records supporting accountability and auditability. Its scope includes agents that initiate or execute financial transactions and other consequential actions on behalf of individuals or institutions. This event is a public speech and an explanation of an industry framework. The accessible SAFR Version 1.0 explicitly states that it does not constitute regulatory guidance or supervisory expectations.

For payment, liquidity management or trading agents, institutions should incorporate authorisation checks into production architecture and procurement acceptance criteria. Translate user authorisation into verifiable action types, amounts and validity periods, and check these before invoking execution systems. Validate each action in a multi-step task independently; block actions that violate hard constraints, refer actions above autonomous execution limits for human approval, and block or escalate approval timeouts under predefined rules. After model or prompt changes, retest unauthorised actions, expired or invalid authorisation, and anomalous inputs. Link audit records to the original mandate, tool calls, applicable rules, approvals and execution outcomes, and restrict access to those records. When procuring external agents, verify support for action interception, identity propagation and evidence export so that they can integrate with institutional control and audit systems. These are deployment recommendations based on the reference framework; specific configurations should reflect business risk and applicable rules.

SAFR was first published as an industry white paper on 3 July 2026. The speech confirmed the release of Version 1.1 and an open-source reference implementation through the Future of Finance Institute (FFI), inviting industry contributions to further development. The speech and the currently accessible white paper can be read alongside the final Guidelines on Artificial Intelligence Risk Management for Financial Institutions issued by MAS on 7 October 2026. The Guidelines establish institution-level expectations for oversight, inventories and life cycle controls, while SAFR provides an implementation reference at the level of individual actions. A further observation point is already established: MAS intends to consult in 2027 on supplementary guidance needed for AI agents in financial services, providing a basis for tracking how runtime practices align with supervisory expectations.

On 7 October 2026, the Office of the Australian Information Commissioner (OAIC) announced a formal privacy investigation into Shenzhen Qingcheng Future Technology Co. Ltd, the developer of the HeyCyan smart glasses phone app. The app is used with Kmart’s Anko smart glasses and other branded products sold through multiple platforms.

The investigation was initiated under section 40(2) of the existing Privacy Act 1988 and concerns the practices, procedures and systems for managing personal information required by Australian Privacy Principle 1 (APP 1). The developer did not respond to preliminary inquiries; a formal investigation enables the use of compulsory information-gathering powers. The examination focuses on how information is recorded, who can access it and where it is stored. OAIC has also urged relevant retailers to consider whether to continue selling the devices. The announcement contains neither a finding of a breach nor a sales ban.

Enterprises procuring or deploying smart glasses should extend due diligence to the apps and data processing providers behind the hardware. OAIC explained on the same day that responsibility depends on who actually holds personal information; selling hardware alone does not automatically create privacy obligations in relation to information collected by the device. Before a pilot, enterprises should therefore verify actual data flows across the “glasses–phone–app–backend service” chain: what stays local, what is uploaded, who can access it and where overseas recipients are located. Contracts should specify vendor responsibilities for providing evidence, responding to inquiries and handling incidents. For on-site collection, assess the ability to notify bystanders, restrict use in sensitive areas and disable recording. Access controls, audit records and privacy risk assessments can form part of production-readiness checks. These are deployment recommendations, rather than new uniform obligations introduced by this investigation.

The action follows preliminary inquiries that began on 12 August 2026, moving questions about a specific developer’s data handling into a formal investigation. The next verifiable milestones are the investigation’s findings and any resulting regulatory decisions; OAIC has stated that it will await the investigation’s conclusion before commenting further. In parallel, on 31 August 2026, the Australian Government released a privacy reform consultation paper and draft legislation proposing a “fair and reasonable” standard for information handling and stronger consent requirements, with explicit attention to AI and devices such as smart glasses. Enterprises should distinguish two governance tracks: data handling responsibilities under existing law are already open to investigation, while reform proposals should be tracked for governance design. Standards that remain proposals should not be presented as obligations already in force through this investigation.

On 9 October 2026, the Opinions on Developing New Quality Productive Forces, issued by the CPC Central Committee and the State Council, were publicly released. As part of the comprehensive implementation of the “AI Plus” initiative, the document explicitly calls for systems for technology monitoring, risk early warning and emergency response.

The Opinions incorporate industrial AI applications and safety safeguards into the same policy programme and constitute a published, comprehensive industrial policy document. Item (10) addresses traditional industry upgrading, smart terminal applications, the supply of computing power, algorithms and data, and pilot-scale testing facilities for sectoral applications and high-value use cases, alongside monitoring, early warning and emergency response mechanisms. Item (15) calls for stronger governance in areas including AI and enhanced joint regulation across departments. The text does not specify enterprise AI risk-classification methods, log-retention periods, incident-reporting deadlines or a standard production-approval procedure.

For projects such as industrial quality inspection and equipment control, enterprises should make safety capabilities part of the acceptance criteria for pilot-scale testing and deployment expansion. Define permitted uses of production data and access permissions at the data stage; test models for missed detections, incorrect judgements and abnormal operating conditions; and distinguish permissions to provide recommendations from permissions to execute actions directly. Require human confirmation and deactivation measures for actions such as changing equipment parameters. Procurement contracts should address model version identification, change notifications and support for safe shutdown or fallback following failures. During operation, record anomalies, response actions and recovery outcomes, and identify accountable owners across business, technology and safety teams. These are deployment recommendations based on the policy direction; control intensity should reflect use-case risk and applicable rules. The National Development and Reform Commission’s official interpretation describes commercial AI adoption as a systems undertaking and explicitly calls for assessing risks relating to technology ethics, data security and industry, providing a policy basis for including safety investment in project budgets and implementation plans.

Monitoring, early warning and emergency response systems were not first proposed this week. The State Council’s Opinions on Further Implementing the “AI Plus” Initiative, published in August 2025, had already called for these systems and addressed risks involving model opacity, hallucinations and algorithmic discrimination. The new document embeds this governance approach in the broader policy for new quality productive forces. The next point to track is whether the laws and regulations, application standards and ethical guidelines called for in the NDRC’s interpretation further specify responsibilities and operational requirements. In the same week, MAS issued its final AI risk management guidelines, directly setting out oversight, inventory, testing and life cycle control arrangements for financial institutions, together with implementation dates. The two documents operate at different institutional levels: China’s Opinions focus on cross-sector development and coordinated governance, while the MAS Guidelines focus on institutional implementation. Cross-regional enterprises should share control capabilities while separately checking applicable instruments, responsibilities and implementation milestones in each jurisdiction.