Autonomous AI systems are rapidly expanding their influence beyond purely software environments, increasingly integrating into physical domains such as warehouses, delivery networks, and public spaces. This significant shift is prompting urgent discussions about whether current AI regulatory frameworks, which have historically focused on online harms, bias, misinformation, and harmful digital content, are adequate to cover systems that operate with tangible, real-world consequences. The inherent risks of embodied AI systems, where failures can directly impact physical infrastructure, property, and human safety, necessitate a re-evaluation of governance approaches.

Experts highlight that embodied AI systems amplify existing risks associated with autonomous software. Dr. Ya-Qin Zhang, founding dean of Tsinghua University’s Institute for AI Industry Research, emphasized that any risk in the digital domain becomes greatly magnified in the physical domain, leading to direct physical consequences. This exposure extends to critical infrastructure, transport systems, drones, logistics networks, and smart grids as AI becomes more deeply embedded in physical operations. Consequently, discussions at recent AI summits, particularly in Singapore, have shifted towards operational safety issues commonly associated with highly regulated sectors like aviation and industrial systems, rather than conventional software regulation. Key concerns include ensuring autonomous systems can operate safely and reliably over extended periods in unpredictable real-world environments, alongside robust operational monitoring and post-deployment assurance.

In response to these evolving challenges, Singapore's Infocomm Media Development Authority (IMDA) has taken a proactive stance, publishing version 1.5 of its Model AI Governance Framework for Agentic AI. This framework provides comprehensive guidance for organizations deploying AI agents capable of planning, making decisions, and executing actions across multiple steps to achieve user-defined goals. These agents can interact dynamically with tools, external systems, and other agents, performing tasks ranging from updating databases and writing files to controlling physical devices and conducting transactions. The framework recommends critical governance measures, including strict access controls, continuous monitoring, and human approval at key stages of deployment, recognizing that not all risks can be anticipated before release due to the dynamic nature of agent interactions with their environment. It also stresses gradual rollouts, continuous monitoring, and further testing even after deployment.

IMDA’s framework delves into four iterative governance areas: upfront risk assessment, human accountability, technical controls, and end-user responsibility. It advises organizations to assess agentic AI use cases based on factors such as data and external system access, autonomy levels, task complexity, the scope and reversibility of agent actions, third-party involvement, and overall system complexity. Practical recommendations include limiting agent access to tools and systems, applying least-privilege permissions, defining standard operating procedures for agent workflows, and establishing mechanisms to take agents offline immediately upon malfunction. Recognizing the impracticality of continuous human review at scale for agentic systems, the framework proposes adapting human oversight to include approval at significant checkpoints, especially for high-stakes, irreversible actions, or outlier behavior. It also addresses risks like automation bias and alert fatigue by recommending auditing oversight through metrics like human override rates and response times, alongside automated real-time monitoring for unexpected behavior. Furthermore, it mandates informing users about an agent's capabilities, data access, and their remaining responsibilities, and emphasizes employee training on human-agent interaction and oversight.

Leading companies are already navigating the complexities of deploying AI in physical and critical environments. Grab, for instance, which is piloting autonomous vehicles and delivery robots in Singapore, relies heavily on extensive simulation, testing in both closed and open courses, and continuous monitoring for its deployment governance. Grab's Chief Technology Officer, Suthen Thomas Paradatheth, highlighted the importance of thoroughly testing in simulation and with a limited number of robots before scaling up, acknowledging the