Singapore issues a consumer advisory on personal AI agents and unveils the ARC Framework
📅 November 7, 2026 • OpenGov Asia
The consumer advisory
Three Singapore agencies have issued a joint consumer advisory about personal AI agents. On November 6 the Cyber Security Agency of Singapore, GovTech and IMDA warned that scammers may try to manipulate an agent into carrying out money transfers or exposing data. The advisory sets out 5 safety steps for users, among them limiting the access an agent has, keeping control of key actions, and understanding how data is used and stored.
The ARC Framework
On November 5 at the STACK Conference 2026 GovTech presented the Agentic Risk and Capability Framework, known as ARC. The framework maps what agents are able to do against the safeguards that should sit next to each capability. Actions with a higher effect, such as financial transactions, official communications or data deletion, require human approval before they are carried out. GovTech describes the approach as applying across the whole of government.
The platform behind the agents
The same speech covered the infrastructure that those agents run on. The Rabbit Deploy platform supports over 8,000 users across 90 agencies, with over 5,000 active applications. One treasury application reduced monthly data consolidation from 12 hours to 1 hour, which adds up to more than 130 staff-hours saved per year for the team involved.
Monitoring and previous work
GovTech also said that GovHarness, shared infrastructure for building and operating AI agents across government, is in development. The name points to a practical need. When many agencies build agents on one platform, the questions of what an agent did, what it touched and who approved the action need consistent answers rather than separate logging conventions in each team. The ARC Framework also extends a line of work that started in January 2026 with the Model AI Governance Framework for Agentic AI from IMDA.
What it means in practice
For observers outside Singapore, the interesting part is that one week produced both a framework for public sector agents and a practical advisory for people using commercial agents. Both documents treat the approval step as the control point and neither presents autonomy as a default. The framework sets expectations for people who build and run agents, while the advisory sets expectations for the people who hand over access to them. The direction of travel is the same in both cases: define the rules for confirmation before agents act at scale. It is also a signal for vendors and platform builders, since a published approval taxonomy is easier to implement than a general call for caution. The open question is how the framework stays current as agent capabilities widen and what will happen when an agent is granted a long chain of approved steps instead of a single one.
IMPACT Singapore is setting approval rules for agents and at the same time reminding consumers to keep control of key actions, and both messages treat a person in the loop as the practical control point for automated tools.
Sources: https://opengovasia.com/singapore-sets-governance-framework-for-agentic-ai-in-public-services/ — https://tech.gov.sg/media/opening-address-by-sms-tan-kiat-how-at-govtech-stack-conference-2026-day-2 — https://channelnewsasia.com/singapore/personal-ai-agents-assistant-advisory-muse-dots-hermes-grok-openclaw-6529836 — https://isomer-user-content.by.gov.sg/36/b21d4369-83bc-4a8c-bbd7-4875c2b07e6c/Using%20Personal%20AI%20Agents%20Safely%20and%20Securely.pdf
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)