ZetrixwebTwo documents, two questions
The Infocomm Media Development Authority released its Model AI Governance Framework for Agentic AI on 22 January 2026, with an update reported in May. It is voluntary and describes itself as a living document. It is concerned with how organizations keep agentic systems accountable: who is responsible, how risks are assessed, and how humans stay meaningfully in control.
The Cyber Security Agency's addendum on securing agentic AI, finalised in June 2026, is a security document. It describes how agentic systems are built, what attackers can do to them, and what controls reduce the risk. The first asks whether the right people are accountable; the second asks whether the system can be abused.
| IMDA framework | CSA addendum | |
|---|---|---|
| Main concern | Accountability and oversight | Security threats and controls |
| Typical readers | Risk, product, legal, senior management | Security, platform, and engineering teams |
| Nature | Voluntary governance framework | Technical guidance alongside 2024 AI security guidelines |
| Useful output | Roles, risk assessment, oversight design | Threat model, control set, test plan |
Using them together
Start with the governance questions: what is this agent allowed to do, who owns the outcome, and what must always need human approval? Those answers set the boundaries. Then use the security guidance to make those boundaries real: scoped credentials, tool restrictions, validation, logging, and testing.
Where the two meet is oversight. A human approval step is only as good as the system that enforces it, so governance decisions should translate into specific technical controls with owners and tests.
Governance sets the rules
Decide permitted actions, accountable owners, and approval points before building.
Security enforces them
Implement the rules as permissions, validation, logging, and adversarial tests.
If you are in financial services
MAS has now published final AI risk management guidelines for financial institutions, which add supervisory expectations on top. For a regulated firm the practical stack is the MAS expectations, the IMDA governance framework, and the CSA security guidance, each answering a different layer of the same question.
Keep a single document that maps your controls to each, rather than three parallel programs.
Rolling out agents in Singapore and need one coherent control map? Talk to us.
Key takeaways
- IMDA's framework is a voluntary governance document about accountability and oversight; CSA's addendum is technical security guidance.
- Governance sets what an agent may do; security makes those limits enforceable.
- Different teams should own each, with a single control map linking them.
- Financial institutions add MAS expectations as a third layer.