technical report

ISO/IEC 27090: the new standard for securing AI systems

In October 2026 the first international standard of the 27000 family dedicated to protecting AI systems against AI-specific attacks was published: data poisoning, model theft, prompt injection. This 4-page note summarises what it covers, what it does not, how it fits with the standards you already know, and the first five steps.

Pages
4
Format
PDF
Language
Español
Cost
Free

What it covers

  • The standard’s complete profile: published 14/10/2026, 56 pages, non-certifiable guidance from the same committee that maintains the 27000 family.
  • Why conventional cybersecurity is not enough: data is part of the code and legitimate access can be the attack vector.
  • The nine threats it covers, from data poisoning to injected outputs, and what each one compromises.
  • How it fits with ISO/IEC 27001, 27002 and 42001: the technical piece the management standards do not develop.
  • What the standard does not do: it is not certifiable, does not address defensive AI and does not guarantee immunity.
  • The first five steps to start without waiting to have the standard in hand.
  • The crossover with personal data protection when models touch people’s data.

Who it is for

  • Information security leads with a 27001 ISMS whose scope already includes (or will include) AI systems.
  • Organisations moving towards ISO/IEC 42001 that need the technical security layer.
  • Development and data teams integrating their own or third-party models into products and processes.