ISO/IEC 27090 is now available after around four years of development: a standard that specifically addresses the security of AI systems. The guide describes attacks aimed specifically at systems with machine learning and generative models and names suitable countermeasures. It is designed as a supplement to ISO 27001, the widely used standard for information security management systems.

The standard thus addresses a problem that has accompanied many companies since the rapid spread of AI functions: established security processes work for classic applications, but they cover the peculiarities of language models and training data only inadequately.

Why AI systems need their own protective measures

Classic applications cleanly separate program code and user input. With language models, this boundary blurs: instructions and content lie in the same text channel. An attack can therefore be hidden in a seemingly harmless document that an AI assistant receives for summarisation.

The best-known case of this kind is so-called prompt injection, i.e. the smuggling in of instructions via regular inputs or processed content. The goal is to move the model to actions that the operator did not intend – such as disclosing internal information or triggering actions in connected systems.

There are also risks along the data chain: training and reference data, models from external sources, interfaces to third-party providers. Each of these stations is a possible attack point that cannot be secured with firewall rules and password policies alone.

Connection to existing structures instead of a parallel world

The decision to attach ISO/IEC 27090 to ISO 27001 is the most important point from an organisational perspective. Companies that already operate an information security management system do not need to build a second governance structure. They can classify AI-specific risks within the existing risk assessment and handle them with the usual mechanisms: identify risks, assess them, define measures, check effectiveness.

For organisations without certification, the standard still has value. It provides a structured list of what can go wrong with AI applications at all – and in practice this list is often the actual hurdle. Many project teams know that they should secure their AI function but have no systematic overview of the types of attack.

What this means for AI functions in websites and CMS

AI components are currently moving into areas that were long considered uncritical. A chat assistant on the website, an automatic text suggestion function in the editorial system, an AI-supported search across internal documents: all three scenarios process content that is not fully under the operator's control.

Three questions are particularly relevant in practice:

  • Which data reaches the model? Content from the CMS, user input, external documents – every source needs a conscious decision as to whether it is trustworthy.
  • What may the model trigger? The more interfaces an AI assistant can serve, the greater the damage in the event of successful manipulation. Tight permissions are the most effective single measure here.
  • How are incidents detected? Without logging of inputs and outputs, it is impossible to reconstruct afterwards what happened.

These questions are not an invention of the new standard, but a standard makes it easier to enforce them in the project. Anyone who can refer to a recognised document does not have to justify security effort anew in every individual case.

A guide, not an automatism

A standard does not replace implementation. ISO/IEC 27090 is designed as a guide and describes attacks and countermeasures – translating them into one's own architecture remains the task of the project team. Especially with generative models, there is also no measure that completely rules out manipulation attempts. A tiered approach to security is realistic: check inputs, limit rights, control outputs, log activity.

The time horizon is equally important. Attack techniques on AI systems continue to develop quickly, while standards, by contrast, emerge over years. ISO/IEC 27090 therefore provides a robust basic framework rather than a conclusive answer to every new attack variant.

Classification for web and software projects

For ongoing projects, this means: AI functions should in future be part of security considerations just like login mechanisms or payment interfaces – with their own set of risks that can now be aligned with a recognised document. Anyone who operates an information security management system can attach the new topics to the existing structure instead of building parallel processes. And anyone who is only planning AI features should consider permissions, data sources and logging from the outset, because retrofitting security in this area is significantly more complex than a clean architecture at the start of the project.

Sources