Singapore financial institutions
MAS AI Risk Management Guidelines: what an AI gateway covers
MAS's Guidelines on Artificial Intelligence Risk Management apply to every financial institution and every form of AI, AI agents included. Inventory and oversight expectations apply from 7 October 2027 and life cycle controls by 7 October 2028. This page maps the parts a gateway can enforce on every model and tool call to what Leanroute does today, including what it doesn't do yet.
Quotes are from the Guidelines as issued on 7 October 2026. Leanroute is not affiliated with MAS, and this page is not legal or compliance advice.
The mapping
AI inventory (Section 4)
Covered“For AI agents, relevant attributes may include agent-specific identifiers, information on the tools and systems the agent can access, its components, and the guardrails imposed to manage risk.”
One API key per agent, named with X-Agent-Name. Each key shows which MCP servers it may reach, its spend limits, and the organization's guardrails, in one dashboard and through the API. How it works
Approved AI tools
Covered“Maintenance of an approved list of AI tools, and a defined process for requesting approval of new AI tools.”
Models and providers are allowed or blocked per organization. MCP servers must be on the organization's allowlist, which only owners and admins can change in the dashboard; each key can be narrowed further. How it works
Monitoring (Section 5)
Covered“Information flow and decision-making paths across workflows that use AI, such as reasoning processes, actions taken, tools used should also be monitored.”
Every model call and tool call is logged and grouped by agent run: model actually served, tokens, cost, status, and each tool with its server, outcome and duration. How it works
Kill switches
Covered“For high risk materiality AI, consider implementing “kill switches” or override mechanisms to rapidly deactivate the AI system if it exceeds risk tolerances.”
Disable one agent's key (reversible) or revoke it, set the organization's spend cap to zero, or cap a single run with a budget that stops it on its own. All take effect on the next request. How it works
Privacy and data leakage
Covered“Privacy risks, e.g., arising from the leakage of confidential or customer data due to the use of third-party Generative AI services.”
Guardrails block or redact personal data and secrets before a prompt reaches a provider, and check responses on the way back. The gateway runs in Singapore and stores no prompt or response content. How it works
Security: prompt injection
Partly“Security risks, e.g., arising from adversarial attacks on Generative AI systems via prompt injection or data poisoning.”
Prompts and every MCP tool result are screened for injection patterns, secrets and personal data before the model sees them; a blocked tool result is withheld and the hit is recorded on the agent run. The checks are pattern-based, and tool arguments are not screened yet. How it works
Enhanced documentation for agents
Partly“Logging prompts and model responses, together with applicable model versions and reasoning processes used.”
Model versions are logged on every call. Prompt and response content is not stored today, by design; optional content logging kept in Singapore is on the roadmap.
Human oversight (Section 5)
Not yet“Mechanisms to escalate AI outputs or recommendations for human review where pre-defined conditions (e.g., on reliability and accuracy) are met.”
No approval step in the gateway yet. Build it in your agent or workflow tool for now; a human-approval hook is on the roadmap.
Change management (Section 5)
Partly“An FI should develop and implement comprehensive and robust controls for managing changes to deployed AI.”
Every change to keys, spend caps, provider keys, guardrails, the MCP allowlist, routing and the team is recorded in the audit log: who, when, and the values before and after, exportable as CSV. Approval before a change takes effect, and rollback, are not built in. How it works
Third-party AI: updates and changes
Covered“The extent to which the FI is able to receive notifications of updates or changes to third-party AI post-deployment, and manage and assess the impact of such updates or changes.”
Every response names the model that actually served it. When a provider retires a model, the call is mapped to its named successor and a response header says so, and changes are published on the changelog. How it works
Third-party AI: contingency plans
Covered“Whether robust contingency plans are in place to address potential failures, unexpected behaviour of third-party AI, or discontinuing of support by vendors.”
One API in front of 14 providers, with failover on upstream errors, so a provider outage or retirement doesn't mean rewriting the integration. Failover can be turned off per request where a use case must stay on one model. How it works
What stays with the institution
Most of the Guidelines are about governance and judgement, which no gateway can do for you:
- Board and senior management oversight, and AI in your risk appetite (Section 3)
- Identifying AI use across the institution and rating each use case's materiality (Section 4)
- Fairness, explainability, testing and pre-deployment review of each use case (Section 5)
- Staff competence, training and resourcing (Section 6)
Leanroute as a third party
Under the Guidelines, Leanroute is third-party AI that your institution onboards and oversees. For your due diligence:
- Operated by Leanroute Pte. Ltd. in Singapore; gateway, database and cache in Singapore
- Named sub-processors and a PDPA-aligned privacy policy
- Monthly automated security scans; latest report on request
- Not yet: SOC 2 report, SSO, or a deployment inside your own environment
Questions
- When do the MAS AI Risk Management Guidelines apply?
- MAS issued them on 7 October 2026. They take effect on 7 October 2027 for oversight, identification, inventory and risk materiality (Sections 3 and 4), and life cycle controls and capability (Sections 5 and 6) are expected by 7 October 2028.
- Do the Guidelines cover AI agents?
- Yes. They apply to all forms of AI, name AI agents and multi-agent systems, list agent attributes for the inventory, and expect monitoring of actions taken and tools used where relevant.
- Does using Leanroute make an institution compliant?
- No tool does. The institution remains responsible, and Leanroute is third-party AI under the Guidelines. A gateway helps by enforcing some controls on every call and keeping records the institution can use as evidence.
- Where does our data go?
- Requests are handled by the gateway in Singapore, which does not store prompt or response content. Each request then goes to the model provider you chose, in that provider's region.
Preparing for October 2027? Tell us which controls matter most to you; the gaps above are what we're building next.