Singapore weighs stricter rules for autonomous AI agents

Moving beyond chatbots

Singapore is currently evaluating how to regulate high-risk artificial intelligence, specifically systems designed to act autonomously. Unlike standard AI that simply answers questions, these 'agent AI' systems can plan and execute complex tasks on behalf of users. This includes performing operations like processing payments or updating customer databases.

The government's interest stems from the fact that these systems can access private data and utilize external tools to interact with the real world. Because they can perform actions rather than just creating content, they pose different risks to business operations. Minister for Digital and Information Development Josephine Teo noted that the government is exploring measures like stricter testing and independent assessments of safety safeguards.

While no new mandatory requirements have been announced yet, the focus is on how AI is deployed and the potential harm it might cause. Policy adjustments will be based on the effectiveness of current controls and the specific risks of the AI's intended use. The goal is to ensure that as AI becomes more capable, it remains within permissible limits and can be audited when problems occur.

Risk-based oversight models

To manage these risks, Singapore's Infocomm Media Development Authority (IMDA) introduced the Agent AI Governance Framework in January 2026. This framework provides guidance on how organizations can limit an AI's autonomy and identify which specific actions require a human to sign off. It encourages businesses to assess risks before deployment and limit the tools the AI is allowed to use.

The framework was updated in May 2026 to include over 10 real-world case studies showing how to categorize AI operations by risk level. For instance, a Singapore-based automation company, Dayos, categorizes tasks based on their impact and reversibility. Low-risk tasks like password resets can be automated and checked later, while high-risk actions like changing access permissions are not permitted to be automated at all.

Software like Tencent's CodeBuddy provides a practical example of these principles in action by defaulting to human approval for sensitive tasks. These include editing files, running system commands, or sending network requests. By requiring a human to intervene in these specific moments, companies can prevent an autonomous system from making a mistake that could compromise the entire network.

Testing in the sandbox

Singapore has been actively testing these control measures through the AI Agents Sandbox program. This initiative began in August 2025 as a joint effort between Google, the Singapore Cyber Security Authority (CSA), GovTech, and IMDA. The four-month test aimed to see how much human oversight is truly necessary for different types of AI tasks.

The results from the sandbox suggested that human approval is not strictly required for every single action an AI takes. For low-risk or remedial operations, the results indicated that testing the outcome after the action is completed may be sufficient. However, the findings emphasized that actions with the potential for serious consequences must require prior human approval before they are executed.

These tests are part of a broader effort to develop methodologies for assessing the security and reliability of autonomous systems. The government is looking at various methods including controlled environment testing and attack simulations. These methods help determine the right balance between allowing AI to be useful and ensuring it does not cause unintended damage to a business or its users.

Cybersecurity and data risks

Increased autonomy in AI introduces specific new cybersecurity threats that regulators are now addressing. One notable example is the 'indirect malicious directive injection' attack. In this scenario, an attacker hides dangerous instructions inside data or content that the AI receives while it is performing a routine task.

Research performed by Google and in Singapore has demonstrated that AI agents can be tricked into performing unintended actions through these injections. In some cases, these tricks have even led to remote code execution, where an attacker gains control over a system. Additionally, the direct processing of personal data by AI agents increases the risk of privacy violations and data leaks.

In response, the CSA issued security guidance for organizations deploying agent AI to mitigate these specific threats. The guidance emphasizes a phased deployment approach where systems are rolled out gradually rather than all at once. This allows organizations to perform data breach assessments and security testing before the AI is given widespread access to company systems.

Sector-specific frameworks

Singapore is also moving toward a model where different industries have their own specific AI management mechanisms. In July 2026, the Monetary Authority of Singapore (MAS) introduced the SAFR framework to govern how agent AI is used within the financial sector. This framework proposes specific measures regarding record keeping, human oversight, and decentralization.

This approach reflects the belief that the level of control should be proportionate to the potential consequences of the AI's actions. For example, a tool that recommends entertainment content does not require the same level of oversight as a system that makes financial decisions for customers. The government wants to ensure that high-stakes industries have the most robust protections.

While these frameworks are being developed, Singapore has not yet set a specific threshold for when mandatory regulations will apply to high-risk AI agents. There is also no confirmed timeframe for when these additional requirements will become law. The current focus remains on developing the tools and guidelines that will eventually form the basis of these mandatory rules.

}

try_itale_it_today_or_empty_string_if_nothing_honest_fits_the_source_mentions_nothing_about_this_but_you_could_check_the_imda_website_for_the_agent_ai_governance_framework_guidelines

watch_questions_or_things_to_watch_next_to_watch_next_[