ISO 27090: Threats to AI Systems

· 4 min read · Lyudmil Arkov

ISO 27090 is a security guide for AI systems. ISO/IEC 27090:2026 was published on 14 October 2026. ISO describes it as guidance for understanding, detecting, and mitigating threats that are specific to AI, across the life cycle. The threats named on the ISO page include data poisoning and model theft. It complements ISO 27001 and ISO 27002. It does not replace them.

Is there an ISO 27090 certificate?

No. ISO 27090 is not a management system standard. There is no accredited ISO 27090 certificate to put in a security questionnaire. If a vendor says they are “ISO 27090 certified,” ask which certificate they mean. The honest answer is a risk assessment that used the guidance, or a confusion with ISO 27001 or ISO 42001.

Use it as an input, not as a control catalogue you paste into a Statement of Applicability. Conventional controls still cover the software, the infrastructure, and access to the data. ISO 27090 is the extra reading for the model and for the data that shapes it, including data collected in operation that is later used to retrain the model.

An ISO page that still says the document is “under publication” is behind the standard page. Cite the standard page for the status. The publication date there is 14 October 2026.

What we will and will not list

Data poisoning and model theft are the examples ISO itself puts on the public page. Secondary write-ups add a longer menu of attack names. We are not copying that menu into a client SoA. Buy the standard if you need the clause titles, and then decide which threats are plausible for the system in front of you.

A coding assistant trained on your own repository is not the same system as a model that accepts public uploads. The first question is still the one ISO 27001 already asks: what asset is it, who can change it, and what happens if the output is wrong or stolen?

How it sits next to the EU AI Act

ISO 27090 is not a harmonised standard. Commission Implementing Decision (EU) 2026/2274 cites EN 18286 and nothing else. Following ISO 27090 is not treated as proof that you meet the Act. High-risk cybersecurity duties, when they apply, still have to be met on the Act’s own dates. Those dates are in the EU AI Act timeline: Chapter III, Sections 1, 2 and 3, except Article 6(5), from 2 December 2027 for Annex III and from 2 August 2028 for Annex I.

ISO 42001 is the management system for AI governance. ISO 27001 is the management system for security. ISO 27090 is guidance you can feed into both risk assessments. A company that already runs an ISMS does not start over. It adds the AI system to the asset list, the supplier list, and the risk register.

The longer ISMS note, including which Annex A controls usually move, is on 27kay.

What to say when a customer asks

Answer the question they asked. “Do you hold ISO 27090?” The answer is that nobody holds it as a certificate. Then say what you did use it for: which systems are in the risk assessment, whether data poisoning and model theft were in scope, and which ISO 27001 or ISO 42001 controls carry the treatment. Offer the statement of applicability and the risk decisions, not a logo.

If they want a presumption under the EU AI Act, ISO 27090 is the wrong document to send. The cited quality-management standard is EN 18286, and only for the Article 17 requirements it covers. Security guidance and a legal presumption are different pieces of paper.

A practical pass

  1. Put every AI system and every AI supplier in the inventory, including the tools staff already use.

  2. For each one, ask where the training data comes from, who can change the model, and whether outputs are treated as facts.

  3. Read ISO 27090 for the threats that match that design. Keep data poisoning and model theft on the list until you have a reason they do not apply. Add others from the standard, not from a slide.

  4. Record the treatment in the risk assessment. If you exclude a threat, write the reason.

  5. Do not open a second management system to hold this. One ISMS, and an AI management system if you are running ISO 42001, is enough structure.

How we can help

We help security and AI teams use one risk register. If the gap is the ISMS, that work sits with the 27kay practice. If the gap is the AI management system, it sits here, under ISO 42001.

Let’s talk - bring the system, not a list of frameworks.