Singapore authorities published three significant AI governance documents between May and June 2026, offering practical guidance on agentic AI, legal responsibility for autonomous agents, and personal data use in generative AI. The coordinated releases from the Infocomm Media Development Authority (IMDA) and the Personal Data Protection Commission (PDPC) signal a determined effort to reduce legal uncertainty as organisations deploy increasingly autonomous AI systems.
Governance framework for agentic AI
The IMDA's Model AI Governance Framework for Agentic AI (Version 1.5) targets organisations deploying agentic AI, whether developed in-house or procured from third parties. It defines agentic AI by its ability to plan, decide, and act independently across multiple steps, and organises recommended practices around four dimensions: assessing and bounding risks upfront, making humans meaningfully accountable, implementing technical controls and lifecycle processes, and enabling end-user responsibility.
The framework identifies five categories of harmful outcomes for organisations to consider:
- Erroneous actions, such as incorrect appointments or flawed code
- Unauthorised actions that exceed permitted scope or lack required human approval
- Biased or unfair actions producing discriminatory outcomes
- Data breaches, including both leakage and wrongful data modification
- Disruption to connected systems during compromise or malfunction
For multi-agent systems, the framework highlights additional systemic risks: agent sprawl (uncontrolled proliferation without centralised management), collaborative failures like miscoordination or collusion, and unpredictable emergent behaviours that cannot be anticipated by testing individual agents. IMDA advises that governance should be calibrated to risk factors affecting impact (domain sensitivity, data access, reversibility of actions) and likelihood (agent autonomy, task complexity, reliance on third-party solutions). Where possible, controls should be structural and system-level rather than prompt-based, with deterministic safeguards preferred for higher-risk actions.
Organisations should define human approval checkpoints for high-risk or irreversible actions and track override rates and response times to audit oversight effectiveness. Testing should cover task execution accuracy, policy adherence, and tool use, with agents gradually rolled out under continuous monitoring and logging. For end-users, IMDA recommends transparency about an agent's range of actions, data access, and escalation channels, along with training on common failure modes like hallucinations and looping errors. The IMDA describes the framework as a "living document" and invites feedback and more case studies.
Legal responsibility for AI agents
IMDA's discussion paper on Legal Responsibility for AI Agents consolidates working group discussions on allocating civil liability when AI agents act autonomously, use tools, interact with third parties, and cause harm. It identifies autonomy, planning, decision-making, and action-taking as the agentic AI features most relevant to liability, and notes that responsibility may need to be considered across a broad value chain: model developers, tooling providers, platform providers, system providers, deployers, end-users, and impacted third parties.
The paper suggests that existing legal frameworks-including contract and negligence-may address many agentic AI scenarios, but highlights challenges in determining fault, causation, foreseeability, and attribution of knowledge or intention across a complex value chain. Contracts can pre-allocate risk, but their usefulness is limited by privity and bargaining power, especially where consumers or impacted third parties are involved. Singapore's current product liability laws do not cover losses arising from AI.
The paper explores both fault-based and strict liability approaches. Fault-based liability faces difficulties in establishing breach, causation, and foreseeability when agents act autonomously and unpredictably. Strict liability could shift complex apportionment disputes away from end-users and third parties, but risks creating unscoped liability and moral hazards. A detailed hypothetical involving a computer-use agent that hacks a cloud provider's servers and causes data leakage and financial harm illustrates how both approaches could apply across the value chain. The paper does not reach firm conclusions and identifies areas for further study, including value-chain responsibility allocation, protections for actors with limited bargaining power, and responsibility for unforeseeable agent actions.
Proposed data protection guidelines for generative AI
The PDPC's Proposed Advisory Guidelines on Use of Personal Data in Generative AI set out expectations for compliance with the Personal Data Protection Act 2012 (PDPA) during development, deployment, and post-deployment stages. The guidelines distinguish three stakeholder roles: model providers, system providers, and system deployers. System deployers bear primary responsibility for ensuring the AI system meets PDPA obligations. Model providers processing personal data for downstream users and system providers developing bespoke AI systems on behalf of deployers may act as data intermediaries.
Model and system providers should document safeguards for downstream stakeholders, while system deployers must map their roles, document responsibilities, and obtain information on upstream safeguards before procurement or deployment. Model providers and system providers are considered organisations when they process personal data for their own AI development, including when model providers use personal data from downstream systems (such as end-user prompts) for model training.
On consent and notifications, the PDPC views general-purpose notifications like "new product development" as insufficient for obtaining consent to use user data for AI training or fine-tuning. Organisations should provide AI-specific notifications covering the types of personal data used, how it will be used, model functions, and withdrawal options. The public consultation on the proposed guidelines closes on 14 July 2026.
Why this matters for legal professionals
These publications directly affect how legal professionals advise clients on AI deployment, data protection compliance, and liability allocation. The agentic AI governance framework provides a practical benchmark for risk management and contractual safeguards. The legal responsibility paper surfaces unresolved liability questions that will shape litigation strategy and regulatory expectations. The PDPC's proposed guidelines demand tighter consent mechanisms and clearer documentation of data flows-areas where legal review is essential. Staying current with these developments is critical, and resources such as AI for Legal can help track evolving governance requirements.
Your membership also unlocks:
