ISO/IEC 42001 · ISO/IEC 27090 · EU AI Act

AI governance and AI security under ISO 42001 and ISO/IEC 27090

We help you use AI securely, transparently and in line with the rules: with an AI management system under ISO/IEC 42001, a threat analysis based on the new ISO/IEC 27090 and clear rules for Copilot, ChatGPT and your own AI agents.

Book an initial consultation Compliance check

What this is about

Why AI governance matters now

ISO/IEC 42001:2023 is the first international standard for AI management systems. It follows the same structure as ISO 27001 and therefore integrates well into an existing ISMS.

Since 14 October 2026, there is also ISO/IEC 27090: the first international guidance against attacks on AI systems. It describes how you detect and defend against threats such as prompt injection, data poisoning or model stealing, and adds the AI perspective to ISO 27001 and ISO 27002.

The EU AI Act has been in force since August 2024 and applies in stages. Since February 2025, certain AI practices have been prohibited and companies must ensure sufficient AI literacy among their employees. Swiss companies are affected if they offer AI systems in the EU or if the outputs of those systems are used there.

Even without regulation: if you introduce Copilot or ChatGPT, you need clean permissions, clear rules for confidential data and an overview of which AI is used where.

42001ISO standard for AI management systems

27090ISO guidance against attacks on AI systems, since October 2026

2.2025First obligations under the EU AI Act, including AI literacy

1 ISMSIntegration with ISO 27001 instead of a parallel structure

ISO/IEC 27090

The attacks the new standard protects against

ISO/IEC 27090:2026 applies to all organisations that develop or use AI. Even if you “only” use Copilot or your own agents, you are affected.

Prompt and output injectionHidden instructions in emails, documents or web pages take over a language model or an agent. Defence: separate inputs, minimise permissions, human approvals.

Data and model poisoningManipulated training or context data changes the behaviour. Defence: check where data comes from, log changes.

EvasionDeliberately altered inputs deceive a model. Defence: robustness tests and monitoring during operation.

Membership inference and model inversionInferences about training data, including personal data. Defence: data minimisation and access control, important for the revFADP (revDSG).

Model stealingReplicating a model through its interface. Defence: rate limits, authentication and monitoring of the API.

Across the entire lifecycleFrom design and development through operation to decommissioning. We anchor the measures in your ISMS and AI management system.

More on this in the article ISO/IEC 27090: The new standard for the security of AI systems.

Approach

How we implement it with you

AI inventory

1–2 weeks

Which AI systems do you use, for what and with which data?

Classification by risk class

Policy & risk

2–3 weeks

AI policy, approval process and risk analysis for your use cases, including threats based on ISO/IEC 27090.

Controls

project-based

Technical and organisational measures, such as permissions and DLP for Copilot.

Management system

ongoing

Integration into your ISMS, monitoring and preparation for ISO 42001 certification.

Outcome

What you have at the end

AI inventoryAll AI systems with purpose, data and risk class.

AI policyClear rules for employees and approvals.

Risk analysisFor your most important use cases.

AI threat analysis based on ISO/IEC 27090Attacks such as prompt injection or data poisoning per use case, with measures.

ISO 42001 gap analysisWith a roadmap to certification readiness.

Copilot readinessPermissions, labels and DLP before the rollout.

AI literacy trainingFor employees and management.

FAQ

Frequently asked questions

Does the EU AI Act apply to Swiss companies?

Yes, if you place AI systems on the market in the EU or their outputs are used in the EU. For purely internal use in Switzerland, data protection and contractual requirements are what mainly apply.

Do we need an AI policy for Copilot?

Yes. Copilot sees everything a user is allowed to see. A policy and cleaned-up permissions prevent confidential data from appearing in answers unintentionally.

Can ISO 42001 be combined with ISO 27001?

Yes. Both follow the same basic structure. We integrate the AI requirements into your existing ISMS and the GRC platform.

What is ISO/IEC 27090?

An ISO guidance standard, published on 14 October 2026. It describes how organisations understand, detect and defend against attacks on AI systems, such as data poisoning, model stealing or prompt injection, and complements ISO 27001 and ISO 27002.

Can you get certified against ISO/IEC 27090?

No. ISO/IEC 27090 is a guidance standard. You provide evidence through certification under ISO 27001 or ISO 42001, in which you implement the measures from ISO/IEC 27090. That is exactly how we build it with you.

How does Cyberopex use AI itself?

Our AI agents handle routine ISMS tasks on the GRC platform, always with a human who approves. We secure them according to the same principles that ISO/IEC 27090 describes.

Where do you stand today?

In a free initial consultation, we clarify your scope and the next three steps.

Book an initial consultation Swiss GRC Hosting