Singapore: IMDA Updates Model AI Governance Framework for Agentic AI and Publishes Discussion Paper on Legal Responsibility for AI Agents

Updated Model AI Governance Framework for Agentic AI

The original MGF was launched in January 2026 as the world’s first comprehensive governance guide specifically designed to address the risks of agentic AI. Following a call for feedback and case studies, IMDA has now refreshed the framework to incorporate industry input from over 60 organisations, including AWS, DBS, Google and Salesforce.

Key changes in the Updated MGF

The Updated MGF retains the four-dimension structure of the original framework — (1) assess and bound the risks upfront; (2) make humans meaningfully accountable; (3) implement technical controls and processes; and (4) enable end-user responsibility — but introduces several notable refinements across each dimension as follows:

• New and expanded risk factors for assessment. The Updated MGF introduces additional risk factors that organisations should consider when assessing agentic AI use cases, including system complexity and the use of third-party solutions. The greater a system’s complexity, the more likely it is to exhibit unpredictable and emergent behaviours arising from unexpected interactions among its components. Additionally, where an organisation relies on agents provided or operated by external parties, it should consider the extent to which its visibility and control over those agents may be limited. The Updated MGF also now expressly includes safety and reliability components (such as controls, logging and monitoring) as part of an agent’s core components.
• Systemic and multi-agent risks. A significant addition is the expanded treatment of systemic and multi-agent risks in risk assessment. The Updated MGF highlights that the speed at which agents take decisions makes it difficult for oversight mechanisms to detect and prevent unauthorised actions in real time before they cause harm. It also warns of cascading or compounding effects: an error at any stage of an agentic workflow may spread and intensify through subsequent stages, leading to disproportionate consequences — for instance, a hallucinated inventory value might trigger downstream processes to order far too much or too little stock. These risks are exacerbated in multi-agent systems, where agents frequently exchange context, memory and intermediate outputs, thereby increasing the potential for sensitive data to be inadvertently logged, transmitted to less secure agents, or compromised through prompt injection attacks.
• Refined agentic AI value chain. The Updated MGF refines its mapping of the agentic AI value chain to more clearly distinguish between the different actors involved in the development, deployment and operation of agentic AI systems. The value chain now identifies distinct roles for platform providers and system providers or application developers, recognising that these actors control different elements of the agentic system and bear different responsibilities. The Updated MGF acknowledges that organisations may play multiple overlapping roles — for example, developing its own agents and subsequently deploying them would render an organisation both the system provider and the deployer. This refined taxonomy supports clearer allocation of accountability.
• Strengthened measures against automation bias. As agents become more capable and perform reliably over time, the tendency for humans to over-trust the system — known as automation bias — becomes a heightened concern. The Updated MGF addresses this by recommending that organisations define significant checkpoints in the agentic workflow that require human approval (particularly for high-stakes or irreversible actions) and regularly audit human oversight to verify that it remains effective. Specific measurable indicators are suggested, including human override rates and human response times during review of agent actions. 
• New categories of technical controls and change management guidance. The Updated MGF introduces an overview of different types of technical controls  (including structural controls, rule-based controls and prompt-layer controls) and provides guidance on how organisations should select among them based on the risks identified. Furthermore, the Updated MGF advises that organisations establish change management and version control procedures. Given that modifications to a single agent within complex agent ecosystems may produce knock-on effects throughout interconnected workflows, organisations should establish clearly defined criteria for triggering change reviews and classify required changes appropriately. 
• Added details on tradecraft and business continuity. The Updated MGF emphasises that over-reliance on AI agents can erode human expertise, which in turn threatens operational resilience in the event of agent malfunction. Accordingly, organisations must actively preserve human skills to ensure continuity by providing sufficient training and maintaining work exposure that retains foundational competencies.

Real-world case studies

A significant addition to the Updated MGF is the inclusion of more than ten real-world case studies from both Singaporean and international organisations, demonstrating how the framework’s recommendations can be operationalised in practice.

Notable examples include Dayos, which implemented tiered risk levels to guide and bound the actions of its AI-powered IT ticketing agent, assessing each ticket type for severity of impact, reversibility and feasibility of human oversight; PwC Singapore, which allocated human accountability across different internal teams for an agentic report-drafting system using three coordinated agents in a primarily sequential workflow; and GovTech Singapore, which adopted a phased approach to rolling out agentic coding assistants within the government, beginning with limited internal use and no external tools before scaling to broader deployment once safeguards such as central logging, monitoring and a framework for connecting to approved external tools were in place.

Discussion Paper on Legal Responsibility for AI Agents

Alongside the Updated MGF, IMDA published the Discussion Paper. The Discussion Paper is the result of a series of deliberations by a committee of subject matter experts convened by the IMDA. It examines how legal responsibility should be allocated when AI agents act autonomously, use tools, interact with third parties and cause harm.

The Discussion Paper is not intended to be a settled position at law; instead, it seeks to develop an initial understanding of the key legal issues and challenges relating to agent liability in private law, and lay out the challenges faced in obtaining redress when agents malfunction.

Key findings and themes

The working group generally agreed that existing legal principles  (including contract law, the tort of negligence, product liability and other statutory regimes)  may be able to address liability issues arising from agentic AI. However, significant practical challenges exist in terms of evidential difficulties, ease of obtaining redress and determining the proper apportionment of liability among actors in the ecosystem. Key difficulties include:

• Difficulties with proving intention or knowledge. Proving intention or knowledge on the part of an AI agent is problematic, which is an essential element under contract law to create legally binding obligations. The challenge becomes particularly acute when an agent acts outside its instructions or behaves unexpectedly. It becomes difficult to attribute the requisite human intention to that specific action.
• Contractual privity limits third-party redress. Third parties impacted by AI use generally cannot enforce contractual protections between other actors in the agentic AI value chain.
• Lack of foreseeability. The unpredictable behaviours of agentic AI systems make it difficult to determine whether harm was reasonably foreseeable, particularly where agents deviate from instructions or produce unexpected outcomes.
• Establishing causation. In multi-agent systems and complex workflows, tracing the causal chain to the ultimate harm is challenging. Where errors cascade through multiple stages or agents interact, determining which actor materially contributed to the loss may be difficult.

Exploring the solution space: fault-based and strict liability

The Discussion Paper explores how two different liability regimes — fault-based (negligence) and strict liability — may operate in an agentic AI system that causes harm. Under a fault-based regime, liability depends on whether an actor failed to take reasonable care. Under a strict liability regime, a party’s fault does not need to be proved; proof of some defect which caused damage or harm suffices.

The fault-based regime maps the various actors in the agentic AI value chain — including model developers, system providers, end users and third parties (such as cloud providers), and considers what each actor controls and what reasonable measures each should take. However, the difficulty lies in the fact that even where all actors have exercised reasonable care, agents may still act in unforeseeable ways, leaving uncertainty as to who should bear the risk of such unforeseen actions. Nevertheless, the reasonable measures outlined provide useful guidance on the level of care organisations should adopt when acting in each role.

Under a strict liability regime, although the end user who suffered harm can be easily compensated, liability apportionment among different actors is unclear. Additionally, concerns are raised that currently, strict liability only applies to inherently dangerous activities and it might be inappropriate to use strict liability for agentic AI, which may deter deployment. Shifting liability away from end users could also introduce moral hazards, disincentivising them from using agentic systems responsibly.

Areas for further study

It is acknowledged that there are still large grey areas and uncertainties around liabilities in agentic AI when they cause harm. The Discussion Paper identifies several areas requiring further examination, including how liability mechanisms may differ as agentic AI becomes more autonomous, how actors with limited bargaining power may be better protected, and responsibility allocation for unforeseeable agent actions.

Practical implications for organisations

Taken together, the Updated MGF and the Discussion Paper signal Singapore’s continued commitment to a practical, progressive and balanced approach to AI governance. Organisations deploying agentic AI should consider:

• reviewing and updating their risk assessments to address multi-agent and systemic risks now covered in the Updated MGF; 
• implementing robust human oversight mechanisms and auditing their effectiveness, including monitoring for automation bias; 
• clearly allocating responsibilities within and outside the organisation across the agentic AI value chain, supported by appropriate contractual arrangements; 
• considering how the civil liability considerations raised in the Discussion Paper may affect contractual risk allocation with third-party AI providers and service recipients; and 
• engaging with the Updated MGF’s real-world case studies to benchmark their own governance practices against industry peers.

Organisations seeking tailored advice on aligning their agentic AI deployments with Singapore’s governance frameworks are encouraged to contact us for further information.