GenAI Assure v2.1: governance and security assurance for AI deployers

The catalogue contains 115 individually testable controls across 17 domains. Each control defines its objective, accountable owner, applicability, evidence requirements and test reference.

Organisations select controls according to architecture, risk tier and actor role, recording the rationale for permitted exclusions.

The framework covers 13 architecture classes: SaaS assistants and copilots; hosted model APIs; retrieval-augmented generation; fine-tuned or adapted third-party models; multimodal systems; workflow orchestration; single agents; multi-agent systems; browser and computer-use agents; plugin and MCP-style integrations; embedded AI; predictive machine learning; and meeting and communication AI.

Ground-up model development and training infrastructure are outside its scope. Where an organisation's activities may create provider obligations, the framework requires specialist escalation.

A seven-stage method addresses prohibited uses, organisational actor status, system classification, mandatory legal overrides, inherent risk, control strength and residual-risk acceptance.

Four risk tiers guide proportionate safeguards. Budget and delivery deadlines do not reduce assessed risk or override mandatory requirements.

The framework addresses prompt injection, sensitive information disclosure, unsafe output handling, overreliance and resource consumption.

It includes RAG source governance and retrieval permissions; agent identities, tool authorisation, containment and interruption; supplier assurance; data minimisation and retention; human review; guardrails and moderation; incident response; and resilience.

Three assurance levels distinguish control design, implementation and operating effectiveness. A policy or configuration record alone does not demonstrate that a control works consistently.

Testing, evaluation, verification and validation requirements support use-case-specific test plans, acceptance criteria, adversarial testing, human-oversight checks and reassessment after material change.

Operational playbooks support shadow AI discovery, incident handling, evidence management and lifecycle governance.

Implementation progresses through Foundation, Control build, Assurance and Advanced profiles. Progress depends on meeting exit criteria and producing sufficient evidence. There is no universal 30–60–90-day completion promise.

The framework includes a crosswalk to instruments such as the EU AI Act, GDPR and UK GDPR, NIST AI RMF and its Generative AI Profile, NIST CSF 2.0, ISO/IEC 42001, OWASP guidance and MITRE ATLAS.

The crosswalk explains where the framework supports alignment and where coverage is limited. It does not establish certification, regulatory endorsement, production readiness or legal compliance.