Abstract

The 2024 wave of AI risk instruments (NIST AI 600-1, ISO/IEC 42001, the EU AI Act) and the CSF v2.0 update establish what an organization must govern, but each is deployment-agnostic and leaves the multi-tenant SaaS engineering layer to the implementer. This paper builds and tests a deployment-context profile for it. Seventeen technical vulnerability classes are recut into five Risk Domains carrying accountable owners, board-visible indicators and regulatory touchpoints, and the lifecycle is decomposed into twenty control families. A multi-tenancy amplification factor, split into an architectural bound and a realisation coefficient, supports a Monte Carlo treatment of platform loss. Evaluation uses 30 publicly disclosed incidents from MITRE ATLAS and vendor advisories, January 2023 to January 2026, with Wilson intervals on proportions and a paired bootstrap over incidents on coverage scores; what each construct claims and what would refute it is fixed before any result is reported. The headline comparison ranks eight existing instruments and excludes our own profile, since we scored all of them: the CSA AI Controls Matrix leads at 4.00 of 4 and COBIT 2019 trails at 0.87. Twenty-nine of the 30 incidents map onto the underlying taxonomy, and vendor and supply-chain control is the most frequently decisive control family at 53%, ahead of output sanitisation at 50%. Our profile reaches 3.87 on the same basis, and third place on the externally sampled subset; what separates it is the governance layer, at 6 of 6 on owner assignment, board indicators and cross-tenant amplification against a best comparator score of 4. Two constructs are reported as untested rather than defended. The assessment questionnaire, scoring code and coded dataset are released so every result can be recomputed or contested.

Statistics