Singapore Reviews Safeguards for High‑Risk Agentic AI

On 7 October 2026, the Singaporean government declared that it would undertake a thorough reassessment of the regulatory safeguards that apply to high‑risk agentic artificial intelligence. The move comes amid growing autonomy of systems that act as agents without direct human supervision. The announcement was made in response to rising international concerns about the capacity of such systems to make decisions with far‑reaching consequences, especially when they are deployed in sensitive sectors such as critical infrastructure and government digital services. The ministry emphasized that the aim is to strengthen the regulatory framework before any real incident occurs, and to define clear criteria for classifying applications as high‑risk. The review will examine existing guidelines, licensing requirements, and oversight mechanisms to ensure they are fit for purpose in the context of rapidly evolving AI capabilities.

What Is Agentic Artificial Intelligence?

Agentic artificial intelligence refers to systems that are capable of taking independent actions in pursuit of defined objectives, without the need for continuous human direction. These systems learn from data and develop their own strategies, which makes them particularly efficient at identifying vulnerabilities and threats within networks and critical infrastructure. Because these agents can autonomously scan networks, they may discover zero‑day vulnerabilities that human operators might miss, but this same capability can be turned against the very systems they are meant to protect if misused. However, the very independence that gives them this efficiency also creates new risks when the systems exceed the boundaries of their original programming or are exploited by unauthorized actors. For this reason, certain uses are classified as “high‑risk” and require close monitoring.

Current Risk Assessment

In a parliamentary response, Digital Development and Information Minister Joselyn Tiew explained that the government has not received any official report of a cyber‑attack carried out by an unsupervised agent. She added that the absence of actual incidents does not mean that the risk is absent; rather, it calls for proactive preparedness. She noted that government agencies already employ artificial intelligence to identify vulnerabilities and threats in their systems, but that clear reporting mechanisms remain necessary to ensure a rapid response should any breach occur. She also highlighted that the current internal reporting channels are being evaluated for speed and clarity, to minimise any lag between detection and response.

Minister Tan Kiatt Hoo, on his part, said that the risks being discussed today are an extension of challenges that have existed for years, and that the government is conducting experiments to develop internal capabilities and a better understanding of responsible AI deployment. He clarified that reviewing regulatory frameworks and incident‑reporting procedures is a fundamental step toward strengthening trust between the public and private sectors, with a focus on building specialised technical teams capable of monitoring the behaviour of intelligent agents and assessing their impact on national security. He further noted that pilot projects are underway to test sandbox environments where high‑risk agents can be evaluated under controlled conditions before any live deployment.

Examples of Use‑Case Risks

During the parliamentary debate, an internal incident was cited in which a civil servant used AI‑generated text to discover more than 95,000 email addresses within a government database. Although the exercise did not result in the leakage of sensitive information, it highlighted the ability of smart tools to extract massive amounts of data at an unexpectedly rapid pace, raising questions about the controls needed to prevent similar exploitation in the future. The incident also prompted a review of data‑access permissions, ensuring that only authorised personnel can trigger large‑scale data extraction using AI tools.

A comparable Australian case was also mentioned, where the use of an unsupervised AI model led to the exposure of sensitive information within a government sector. Sources did not disclose further details, but the example demonstrated that the risks are not confined to a single geography and can recur in any nation that relies on similar technologies without stringent supervision mechanisms. The Australian example underscored the importance of cross‑border collaboration on AI safety standards, as vulnerabilities in one jurisdiction can have ripple effects elsewhere.

Proposed Measures and Framework Update

The government used these facts to outline a set of safeguards that will be incorporated into the new regulatory framework. Proposed measures include strengthening internal capabilities to assess the behaviour of intelligent agents, developing unified reporting protocols that obligate all public and private entities to disclose any potential incident immediately, and introducing requirements for risk assessments prior to the deployment of any high‑risk agentic system, with periodic reviews to ensure alignment with international standards. The framework will also require documentation of the decision‑making logic of agents, enabling auditors to trace how specific actions were derived from the underlying algorithms.

• Establish a dedicated unit within the Ministry of Digital Development to monitor the behaviour of agentic systems.
• Mandate periodic incident reports from all entities that use high‑risk AI models.
• Require AI model providers to implement rapid‑response reporting mechanisms when unexpected behaviour is detected.
• Conduct pre‑deployment risk assessments before any agentic system is introduced into critical infrastructure.

In closing, Minister Joselyn Tiew affirmed that the review of safeguards will be completed within the next quarter, with a draft of the regulatory rules slated for public consultation before the end of the year. Consultation sessions will be held with the private sector and the research community to ensure the framework is comprehensive and effectively implemented. Stakeholders will be invited to submit feedback on draft provisions related to transparency, accountability, and penalties for non‑compliance, ensuring that the final rules reflect a balanced approach between security and innovation. The government stresses that the objective is not to slow innovation, but to steer it toward a responsible path that guarantees cybersecurity and economic stability for the country.