airmkit

Operationalise the MAS Guidelines on AI Risk Management for LLM and agentic AI systems.

The Monetary Authority of Singapore issued its Guidelines on Artificial Intelligence Risk Management on 7 October 2026. They take effect on 7 October 2027 for oversight, identification, inventory and risk materiality assessment, and on 7 October 2028 for life cycle controls and capability (§1.8, as clarified by the Response to Feedback §13.5).

The Guidelines are 30 pages of principles. airmkit turns them into artefacts an engineer can produce and an auditor can read: a machine-readable clause index, an AI inventory schema with first-class support for agents, a transparent risk-materiality rubric, a control catalogue where every control cites a paragraph, evidence records that cannot be quietly edited, and an audit pack.

Read the paper on the web · PDF · Docs · One page for a CRO

Not affiliated with MAS. Not legal or regulatory advice. Independent open-source project. Every interpretation and every threshold is the authors'; MAS declines to prescribe materiality formulae, metrics or testing methodologies (Response to Feedback §7.19, §10.28, §10.66). See DISCLAIMER.md.

Why

Most AI governance tooling stops at a spreadsheet of policies. The hard part of these Guidelines is not writing the policy, it is answering three questions per use case, repeatedly, with evidence:

1. What AI do we actually have? Including AI embedded in SaaS the vendor does not call AI (§4.2 fn21), and shadow AI staff adopted without asking (§4.4 fn22).
2. How material is it? On Impact, Complexity and Reliance (§4.12), where moving a use case from "staff reviews every suggestion" to "the agent calls the tool itself" changes the answer.
3. Which controls bind, and can we show they work? Including the expectation that guardrails be tested for effectiveness, not merely present (§5.15 fn46).

airmkit answers all three from files in your repository, and prints the paragraph behind every answer.

Install

pip install airmkit                 # core
pip install "airmkit[evals]"        # adapters for inspect-ai, deepeval, ragas, garak
pip install "airmkit[report,docs]"  # DOCX output and the docs site

Python 3.11+. No GPU, no API key and no network access required for the core or for make reproduce.

Three-command quickstart

airm init --institution "Merlion Commercial Bank"   # scaffold a register and rubric
airm scan . --out candidates.yaml                   # find AI already in your code and configs
airm report --register ai-inventory.yaml --out audit-pack/

The third command writes the audit pack: an inventory register, a materiality report, a control-applicability statement, a gap analysis by effective date and a one-page board summary.

$ airm assess --register examples/synthetic_bank/ai-inventory.yaml

  Use case                     Impact      Complexity  Reliance    Inherent   Residual
 ─────────────────────────────────────────────────────────────────────────────────────────
  credit-decision-model        critical    medium      high        Critical   High
  aml-alert-triage             high        medium      medium      High       Medium
  kyc-document-agent           high        critical    high        Critical   High
  policy-rag-assistant         low         medium      low         Medium     Low
  ...
  Rubric: default-v1 · dimensions MAS-AIRM §4.12(a)-(c) · tiering is this toolkit's proposal

$ airm check --register examples/synthetic_bank/ai-inventory.yaml --use-case kyc-document-agent

  FAIL  AIRM-5.15-GUARDRAIL-EFFECTIVENESS   guardrail 'injection filter' has no effectiveness test
        MAS-AIRM §5.15, §5.15 fn46 · auto
  ATTEST AIRM-5.19-INDEPENDENT-VALIDATION    validator must be independent of dev and deployment
        MAS-AIRM §5.19 · semi · inherent tier Critical
  PASS  AIRM-4.8-FN25-AGENT-ATTRIBUTES      agent identifier, tools, components, guardrails present
        MAS-AIRM §4.8 fn25 · auto

What is in the box

JSON Schemas for every model live in schemas/, so a GRC platform or a spreadsheet can adopt the formats without Python — regenerate them with airm schemas.

Full documentation, including a concepts guide, the CLI reference and a cookbook, is at https://thanatchon36.github.io/airmkit/ (make docs builds it locally).

The paper

Operationalising the MAS Guidelines on AI Risk Management for LLM and Agentic Systems: A Machine-Checkable Control Catalogue and Empirical Study.

Research

The repository is also a reproducible study of how far these Guidelines can be machine-checked. Five questions, thresholds fixed in advance, in docs/RESEARCH_QUESTIONS.md: automatability of each expectation, how autonomy changes materiality, measured guardrail failure rates on a local model, crosswalk coverage, and the shape of the compliance workload.

Of 17 pre-registered hypotheses, 10 held and 7 were refuted — including that the guardrail patterns which suppress attacks also refuse half of benign in-scope requests, and that what an ISO 42001 + NIST-aligned institution still owes is governance machinery rather than the agent clauses. All of them are reported, with the thresholds they failed.

make reproduce     # rebuilds every table and figure in paper/paper.md
make quick         # under 5 minutes, no model calls

Results are in results/summary.json, the figures in results/figures/ with their captions, and the write-up in paper/paper.md (PDF, web). make paper rebuilds all three.

Clause index

Every paragraph of the Guidelines maps to the module, control and CLI command that implements it. The table is generated from the catalogue, never hand-maintained:

airm clauses --format markdown        # or: make docs

Development

make setup      # uv venv -p 3.11 and an editable install with every extra
make quick      # lint, fast tests and the worked example — under 5 minutes
make test lint  # full suite with coverage; ruff + mypy --strict
make help       # every target

Generated files (schemas/, results/summary.json, paper/paper.md, docs/EXEC_SUMMARY.md, docs/posts/*.md, docs/clause-index.md) have staleness tests: edit the source, then run make schemas paper docs.

Contributing

The most useful contributions are new crosswalks (HKMA, BNM, Bank of Thailand, APRA), new eval adapters, and challenges to the interpretations in docs/INTERPRETATIONS.md — if you read a paragraph differently, open an issue and say which paragraph. See CONTRIBUTING.md.

Citation

@software{airmkit,
  author  = {Chongmankhong, Thanatchon},
  title   = {airmkit: Operationalising the MAS Guidelines on AI Risk Management
             for LLM and Agentic Systems},
  year    = {2026},
  url     = {https://github.com/thanatchon36/airmkit},
  version = {0.1.0}
}

See CITATION.cff. Release notes are in CHANGELOG.md.

Licence

Apache-2.0. The MAS documents reproduced under docs/sources/ remain MAS copyright and are included only so that every citation in this repository can be verified offline.