The Monetary Authority of Singapore has published Version 1.1 of its white paper on Safeguards for Agentic Finance at Runtime, or SAFR, a framework for checking what AI agents do before their financial actions execute. The paper’s version table dates it 9 October 2026, three months after Version 1.0 of 3 July.
MAS Managing Director Chia Der Jiun announced it the same day, in a keynote at the INSEAD Digital Finance & Agentic AI Summit. The acknowledgement list grows from eight organisations in Version 1.0 to eleven, by our count, with DBS, Google and UOB added.
The version table gives the reason for the update in one line: “Clarify scope and implementation approach; Additional case studies”. Version 1.0 is listed as “First publication”.
In his speech, Chia said: “We have recently published Version 1.1 of the SAFR white paper, with contributions from an expanded group of financial institutions and frontier AI developers.”
He also said MAS had released code alongside it: “Through the Future of Finance Institute, we have also released an open-source reference implementation of SAFR to support industry experimentation, implementation and further development.”
The paper is explicit about its status. It “does not constitute regulatory guidance or supervisory expectations”, and MAS’s publication page says SAFR “is jointly developed by the financialindustry under MAS’ BuildFin.ai initiative.”
The eleven organisations named in the acknowledgement list of MAS’s SAFR white paper, Version 1.1 (October 2026); highlighted are the three not named in Version 1.0’s list (July 2026). The count and the comparison are ours.
Business Pill · HUMAN ON THE LOOP
A short explainer of the human-on-the-loop idea: people oversee and step in on exceptions rather than approving every action. It teaches the general idea only and says nothing about any organisation in this story.
The key insight: As we read it, SAFR moves the check from after the fact to the moment before an agent acts: every consequential action is declared, matched to a registered agent and its mandate, and given an outcome before it executes. Version 1.1 widens who stands behind that design, from eight named organisations to eleven, without turning it into a rule: the paper says it is not regulatory guidance.
Who Is Named
Version 1.0’s acknowledgement list names Ant International, Circle, HSBC, J.P. Morgan Chase, Manulife, Mastercard, OCBC and Visa. Version 1.1’s list adds DBS, Google and UOB, for eleven organisations.
Chia’s speech refers to “frontier AI developers” in the plural. Of the organisations on the paper’s list, Google is the one that builds frontier AI models; the list names no other AI developer.
The chart below shows the eleven names, with the three additions highlighted.
How SAFR Checks an Agent
In the speech, Chia summarised SAFR as “three main elements: establishing agent identity and authorisation; checking proposed agent actions against the user’s mandate before execution; and maintaining accountability and auditability.”
The paper itself describes four runtime components. It says SAFR “introduces a governance checkpoint between material agent decision and its execution”, made up of Agent Identity, a Controls Repository, a Disposition Engine and an Audit Log, which pass a single record called the Governance Envelope.
The Disposition Engine gives every proposed action one of four outcomes: Deny, Escalate, Auto-Execute or Observe. Escalate means the action “is held pending human review before proceeding.” The paper lists the factors that set the outcome, including action reversibility, financial materiality and customer impact severity.
Two lines set the limits. “An agent cannot extend the scope of a mandate through its own reasoning or inference,” the paper says, and “The governance envelope should not be treated as self-verifying.”
The four outcomes the Disposition Engine gives every proposed agent action, as defined in MAS’s SAFR white paper, Version 1.1 (October 2026). Highlighted: Escalate, which holds the action for human review.
What Version 1.1 Adds
Comparing the two versions, we found a new passage on how institutions can build it: “SAFR does not require implementation as a single central platform within an institution.” The paper says its components may be implemented centrally, inside individual agent deployments, as shared services, or as a hybrid.
Version 1.1 also adds a risk and compliance case study, in which “AI agents investigate flagged transactions and propose whether a suspicious transaction report (STR) should be filed.” In it, “Proposals to file an STR are escalated to a compliance officer, who approves the filing.” The case names no institution.
The case studies carried over from Version 1.0 cover Ant International’s treasury agents, Mastercard Agent Pay, Visa Intelligent Commerce, Circle’s Agent Wallet, an OCBC project and a Manulife insurance case.
The Structural Read
The paper and the speech describe the same design at two levels of detail. Chia’s summary has three elements: identity and authorisation, checking actions against the user’s mandate before execution, and accountability.
The paper breaks this into four runtime components, Agent Identity, a Controls Repository, a Disposition Engine and an Audit Log, joined by a Governance Envelope that records what the agent proposes and the steps behind it.
Human review is one of four outcomes, not a default. The paper says actions above the threshold for autonomous execution are held for review, and it asks institutions to size escalation to what reviewers can actually process.
MAS, Safeguards for Agentic Finance at Runtime, Version 1.1 (October 2026)
“An agent cannot extend the scope of a mandate through its own reasoning or inference.”
Three Implications
BANKS DEPLOYING AGENTS The paper recommends what it calls native integration for new agents and a gateway pattern for legacy or third-party agents, and version 1.1 says SAFR need not run as a single central platform.
COMPLIANCE TEAMS Version 1.1’s new case has agents propose whether to file a suspicious transaction report, with each proposal to file escalated to a compliance officer who approves it.
PAYMENT NETWORKS The case studies cover Mastercard Agent Pay, Visa Intelligent Commerce and Circle’s Agent Wallet, each mapped by the paper onto SAFR’s four components.
The analysis puts it this way: “The governance layer sits at the intersection of capability, regulation, national security, access control, and geopolitical alignment.”
As we read it, SAFR is a regulator and its banks sketching that layer for one sector: who an agent is, what it may do, and what is recorded, settled before the money moves.
What Is Not Established
The paper does not say whether any institution runs SAFR in production, and it gives no adoption figures or timeline. It is not regulatory guidance, so it creates no obligation for the banks that MAS supervises. Neither the speech nor the paper says what each newly named organisation contributed, or names any AI developer other than Google. We did not review the open-source reference implementation the speech mentions; the paper itself does not describe it.
Business Engineer Framework
Inside Anthropic’s Permission Layer
A Business Engineer analysis of the governance layer of the AI stack: the rules, permissions and boundaries around frontier intelligence.
MAS has revised its framework for checking AI agents’ financial actions three months after the first edition, widening the acknowledged group from eight organisations to eleven, adding a suspicious-transaction case and stating that the controls need not run on one central platform.
The paper describes itself as a starting point for industry discussion, and it invites financial institutions to contribute pilot findings and gaps in the specification.
A note on sourcing. On 12 October 2026 we read in full MAS’s SAFR white paper Version 1.1 and Version 1.0, its publication page and Chia Der Jiun’s keynote of 9 October, linked below. The dates, components, outcomes and case studies come from the paper; the three-element summary, the frontier-developer remark and the reference implementation come from the speech. The organisation counts and the comparison between the two versions are ours, made from the two acknowledgement lists and texts. We did not review the open-source code. Nothing here is investment or regulatory advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
