Safeguards for Agentic Finance at Runtime (SAFR) is an industry white paper published by MAS and financial institutions on 3 July 2026 under the BuildFin.ai initiative, with version 1.1 on 9 October 2026. It proposes a runtime governance layer that sits between an AI agent and the systems it acts on: every consequential action is packaged in a governance envelope, bound to a verified agent identity, checked against a controls repository, resolved by a disposition engine to Deny, Escalate, Auto-Execute or Observe, and written to a tamper-evident audit log. The paper says it 'does not constitute regulatory guidance or supervisory expectations'; it builds on the MindForge AI Risk Management toolkit.
OFFICIAL TEXT: mas.gov.sg ↗ · FINAL · MAS
| Document | MAS SAFR white paper on AI agents — Safeguards for Agentic Finance at Runtime (SAFR) |
| Issued by | Monetary Authority of Singapore |
| Type | Report |
| Status | Final |
| Published | Jul 3, 2026 |
| Applies to | Financial institutions deploying AI agents that initiate actions — payments, trading, credit, claims, regulatory filings — including banks; an industry white paper developed under MAS's BuildFin.ai initiative that 'does not constitute regulatory guidance or supervisory expectations' |
| Official source | mas.gov.sg ↗ |
| Use cases | Generative & agentic AI · AI governance (general) · Model risk management · Fraud detection |
What are the key points of MAS SAFR white paper on AI agents?
- Four runtime components: Agent Identity (binds each proposed action to a registered agent), Controls Repository (the institution's configurable rulebook of policy, regulatory, product and mandate controls), Disposition Engine and Audit Log ('a tamper-evident, append-only record of every governance decision').
- The governance envelope packages the action, the action trace (tool calls, data retrieved, checks performed) and context metadata; the paper warns it 'should not be treated as self-verifying' and should be checked against records captured independently of the agent.
- Four outcomes for every action: Deny (violates a hard constraint), Escalate (held for human review), Auto-Execute (within defined risk thresholds) and Observe (proceeds but flagged for review).
- It names three gaps in existing controls: no pre-execution assurance (model risk management validates before deployment, audit reviews after execution), no consistent human-agent governance interface, and fragmented, non-interoperable guardrails.
- Human escalation must be substantive: institutions should size escalation volume to reviewer capacity, set timeouts for review, and give reviewers 'clear authority to approve, modify, or decline escalated actions'.
- Version 1.1 (9 October 2026) clarified scope and implementation and added case studies from firms including Ant International, Mastercard, Visa, Circle, OCBC and Manulife.
- MAS's 3 July 2026 release cites use cases in agent-assisted payments and treasury, wealth-management and advisory workflows, and client engagement, and invites firms to join the BuildFin.ai work group; reference code is available on request.
What did MAS SAFR white paper on AI agents change for banks?
SAFR is the first Singapore document to specify how agentic-AI controls operate at the moment an agent acts, rather than at model approval or after the fact. MAS's final Guidelines on AI Risk Management (7 October 2026) set the supervisory expectations for AI agents; SAFR is a voluntary reference architecture banks can use to implement the human-oversight, auditability and third-party controls those Guidelines expect.
Is the MAS SAFR white paper a regulatory requirement?
No. The paper says it 'does not constitute regulatory guidance or supervisory expectations, nor does it prescribe or anticipate future directions for such guidance'. It is an industry reference approach developed under MAS's BuildFin.ai initiative. Supervisory expectations for AI agents are in MAS's Guidelines on Artificial Intelligence Risk Management.
What are the four SAFR outcomes for an AI agent's action?
Deny, Escalate, Auto-Execute and Observe. A disposition engine checks each proposed action against the institution's controls and assigns one outcome, and every outcome is recorded in a tamper-evident audit log.
How does SAFR relate to Project MindForge?
MAS says SAFR builds on the MindForge AI Risk Management toolkit, focusing on how safeguards can be operationalised at the point of action for AI agents, with policy-bound execution, real-time validation, auditability and interoperability.
| Date | Document | Status |
|---|---|---|
| Oct 7, 2026 | MAS Guidelines on Artificial Intelligence Risk Management — Guidelines on Artificial Intelligence Risk Management for Financial Institutions | Final |
| Mar 20, 2026 | MAS MindForge AI Risk Management Toolkit — AI Risk Management: Operationalisation Handbook (Project MindForge AI Risk Management Toolkit) | Final |
| Mar 6, 2026 | MAS third-party risk management consultation (P004-2026) — Consultation Paper on Proposed Guidelines on Third-Party Risk Management | Proposed |
| Nov 13, 2025 | MAS Consultation Paper P017-2025 (AI Risk Management Guidelines) — Consultation Paper on Guidelines on Artificial Intelligence Risk Management | Superseded |
| Sep 18, 2025 | MAS Circular MAS/TCRS/2025/06 — Cyber Risks Associated with Deepfakes (Information Paper) | In force |
| Dec 5, 2024 | MAS AI Model Risk Management information paper — Artificial Intelligence Model Risk Management: Observations from a Thematic Review (Information Paper) | Final |
Follow every document these regulators publish
when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning