Three AI governance standards landed in one week. One of them matters more than the other two.

In the week of 14 October 2026, three AI governance standards were published or took effect. ISO/IEC 27090:2026, on AI cybersecurity, was published on 14 October. EN 18286:2026, on quality management for the EU AI Act, was cited in the Official Journal of the European Union in October, after being published in July, which makes it the first harmonised standard supporting the AI Act. IEEE 2863-2026, on organisational AI governance, was published the same week.

Three standards in one week sounds like a lot to absorb. It is not. Two of them are useful reference material. One of them changes what you can prove.

ISO/IEC 27090: security threats that are specific to AI

What it is. Guidance on security threats that only exist because a system uses AI: data poisoning, model theft, and similar attacks, considered across the whole lifecycle from training data to production.

Who it applies to. Anyone running models in production, whether you trained them yourself or call one through an API.

The one thing to do. Check that your security review covers AI-specific threats, not just infrastructure. A clean penetration test of your servers says nothing about whether someone can poison the data your model learns from or extract the model itself.

IEEE 2863: a structure for governance

What it is. A recommended practice describing the governance activities an organisation should establish: principles, oversight, and the processes that put both into practice.

Who it applies to. Anyone standing up AI governance from scratch, which describes most startups and scale-ups.

The one thing to do. If you have no governance structure, use it as one. It saves you designing a framework from a blank page, and it gives you something recognisable to point at when a buyer asks how AI is governed in your company.

EN 18286: the one that changes legal standing

What it is. A European standard for the quality management system that the EU AI Act requires of providers of high-risk AI systems under Article 17.

Who it applies to. Providers of high-risk AI systems, meaning those in categories such as hiring, credit, education and critical infrastructure.

Why it matters more than the other two. It is a harmonised standard cited in the Official Journal. That gives it legal effect the others do not have. A provider of a high-risk AI system that follows EN 18286 gains a presumption of conformity with the Article 17 quality management system requirements.

In plain terms, Article 17 tells you to have a quality management system and leaves a lot of room for interpretation. EN 18286 turns that vague obligation into a checklist. It also gives your buyers' diligence teams a simple yes-or-no question to ask: do you follow EN 18286? Expect to hear it in security questionnaires and procurement reviews.

The one thing to do. Find out whether any of your systems fall in a high-risk category. If they do, treat EN 18286 as your route to Article 17 compliance, and start before the 2 December 2027 deadline for Annex III high-risk systems. A quality management system takes months to build and evidence, not weeks.

Why the pattern matters

For several years, AI governance work was mostly about defining what responsible AI means: principles, values, frameworks. This week shows the shift that has been under way. The new standards specify how to demonstrate it.

That distinction matters because demonstration is what other people will ask you for. Buyers want evidence before they sign. Insurers want it before they price cover. Investors want it in the data room. A statement that you take AI seriously no longer closes any of those conversations. A documented position mapped to a named standard does.

If you are not sure where your systems sit, the [free risk check](/risk-check) gives you a first read in five questions.