AI governance, risk and security assessments

Find out whether your AI program could explain itself under questioning. We review the AI you build, buy and run against the frameworks regulators, customers and boards now ask about.

Schedule a consultationRead case studies

AI risk is not a subset of security risk

A model can be hosted securely, patched and encrypted and still produce outcomes that harm the people it decides about, drift away from the conditions it was validated under, or depend on a third-party model nobody in the organization can explain. No security control catalog surfaces any of that.

Every AI assessment starts with an inventory of what you actually run: the use cases in production and in development, the foundation models and third-party APIs they depend on, the decisions they influence and who is affected downstream. Shadow AI use and embedded vendor features rarely appear on an existing register.

From there we work the chosen framework requirement by requirement. Each one carries a statement of what the organization actually does, a risk rating with the reasoning behind it, the evidence that supports the position and a corrective action where it does not hold.

AI frameworks we assess against

NIST AI Risk Management Framework (AI RMF 1.0)

An outcome-by-outcome review of the AI RMF Core as published in NIST AI 100-1: all 72 subcategories across 19 categories and the GOVERN, MAP, MEASURE and MANAGE functions. The framework is voluntary and has no certification scheme, so the deliverable is a position you can defend.

  • AI system inventory and use-case boundaries
  • The seven trustworthiness characteristics examined separately, not as one score
  • A roadmap sequenced by exposure, not by subcategory number

NIST AI 600-1 Generative AI Profile

An action-by-action review of your generative AI systems against the profile: all 212 suggested actions across 49 subcategories, 19 categories and the four AI RMF functions.

  • Generative AI system inventory, including third-party models
  • Risk rated against the twelve generative AI risks the profile names
  • AI actor interviews backed by technical evidence

ISO/IEC 42001:2023 AI management systems

A requirement-by-requirement review of your AI management system against the first management system standard for artificial intelligence: the 32 requirements of Clauses 4 to 10 and the 38 Annex A reference controls.

  • Scope, AI system inventory and role determination
  • AI risk assessed as consequences to individuals, groups and societies
  • Readiness for a certification audit

ISO/IEC 23894:2023 AI risk management

A review of how your organization manages AI risk against ISO/IEC 23894:2023 and the ISO 31000:2018 clauses it extends, covering 131 requirements across 20 clause groups.

  • Who owns the risk of each AI system, and what happens if one drifts
  • Impact assessment treated as three assessments, not one

ISO/IEC 38507:2022 governance of AI

Adopting AI does not move accountability anywhere. We review how your governing body directs, oversees and reports on the use of AI against 145 requirements drawn from Clauses 4 to 6 and Annex A.

  • Accountability tested where AI makes the decision
  • Scoped to the AI you develop, buy or use

European Union Artificial Intelligence Act

The AI Act is directly applicable law that reaches providers outside the Union whose systems or outputs are used inside it. We work Regulation (EU) 2024/1689 article by article: 316 operator-facing requirements across 19 chapter and section families.

  • Operator role and applicability: provider, deployer, importer, distributor or general-purpose AI model provider
  • Article 5 prohibited practice screening for every system
  • High-risk classification under Article 6 and Annex III
  • Penalty-weighted findings and a sequenced remediation roadmap

OECD AI Principles

The OECD AI Principles are not binding, which is exactly why organizations cannot say where they stand against them. We review the operative text of OECD/LEGAL/0449, as revised in 2024, provision by provision: 86 obligations across 17 families.

MITRE ATLAS threat assessment

A threat assessment of an AI or machine learning system against MITRE ATLAS, the catalog of adversarial tactics and techniques used against AI systems. Techniques are scoped to the platforms you run, scored for likelihood, impact and control effectiveness, and mapped to mitigations.

  • Residual risk by tactic and technique
  • Mitigation coverage and a gap and remediation plan

CIS MCP Server Benchmark v1.0.0

A readiness and gap review of your Model Context Protocol estate against all 55 recommendations in 10 sections, reported per recommendation with a Level 1 and Level 2 breakdown. It covers local stdio and remote Streamable HTTP servers and the host applications that connect to them.

  • Evidence from on-the-wire probes, server audit logs and design records
  • AWS evidence for hosting, gateway, identity, egress and supply chain

Questions about AI governance, risk and security

Which AI framework should we start with?

It depends on who is asking. The NIST AI RMF is the usual starting point for a US organization that needs a defensible position, and NIST AI 600-1 extends it to generative AI. ISO/IEC 42001 fits when customers expect a certifiable management system. The EU AI Act applies as law if your systems or their outputs are used in the Union. We help you choose during scoping.

Can AI governance be assessed from cloud configuration?

Only in part, and we say exactly which part. In the NIST AI RMF, 9 of the 72 subcategories are met through organizational process alone, and in the EU AI Act 215 of the 316 requirements are met through legal determination, documentation or a filing. We name the requirements no cloud export can answer and collect interview and documentary evidence for them.

Do you assess the security of AI agents and MCP servers?

Yes. The CIS MCP Server Benchmark review covers Model Context Protocol servers and the host applications that connect to them, and a MITRE ATLAS threat assessment covers adversarial techniques against AI and machine learning systems.

Can TrueMark also build the AI systems it assesses?

Yes. Our engineers build generative AI and machine learning workloads on Amazon Bedrock and Amazon SageMaker, so remediation of what an assessment finds can be carried out by the same team.

Related AWS services

Talk to an AWS engineer about your environment

Book a consultation with our team in Salt Lake City, Bordeaux or Johannesburg and tell us what you are trying to solve.

Schedule a consultation