How SAFR works
The Monetary Authority of Singapore has published an updated industry framework for checking financial AI agents before they carry out consequential actions. Version 1.1 of Safeguards for Agentic Finance at Runtime, or SAFR, is dated October 9 and adds implementation clarification and case studies. [1]
The framework places a checkpoint between an agent’s proposed action and its execution. It describes verifying the agent’s identity, checking the action against the institution’s controls and the user’s instructions, deciding whether it can proceed, and retaining an audit record. The paper identifies payments, trading, credit, regulatory reporting and insurance claims as relevant uses. [1]
The paper also cautions that an agent’s own account of what it did is not self-verifying. Where feasible, the proposed action and supporting record should be checked against information captured independently, such as system records of tool calls and data access. [1]
SAFR is an industry reference approach. The paper explicitly says it does not constitute regulatory guidance or supervisory expectations; institutions remain responsible for meeting applicable requirements. [1]
In an October 9 speech at INSEAD, MAS Managing Director Chia Der Jiun said an open-source reference implementation had also been released through the Future of Finance Institute to support experimentation and further development. [2]