ISO & IEC / Guidance
ISO/IEC 27090:2026 Protect AI systems from compromise
ISO/IEC 27090 considers security threats to AI systems across their life cycle, including attacks on data, models and their use. It helps teams connect possible compromise to detection and mitigation choices while retaining ordinary cybersecurity foundations.
The place in the system
Where it fits
It complements 27001 and 27002 security practices; 42001 addresses the wider organizational management of AI.
Practical interpretation
How to use this reference
These are editorial starting points, not copied requirements or an implementation checklist from the standard.
Questions and actions
- Inventory AI systems, data sources, model dependencies and update paths.
- Threat-model ways that inputs, training data or access could change model behaviour or expose assets.
- Assign monitoring and response owners for suspicious model use and compromised data.
Evidence to look for
- AI system and dependency inventory
- AI-specific threat review
- Monitoring and incident-response decision record
Illustrative example
What it could look like
A team plans to retrain an internal model using operational data collected from users. It checks how that data could be manipulated, who detects unusual behaviour and how a suspect model version would be contained.
Boundary to keep clear
This is guidance on protecting AI systems, not certification requirements, a guarantee against attacks or a new control added to 27001 Annex A.
Source and edition check
Open the official record
Research checked 2026-10-20. Catalogue status and edition can change; the official publisher record controls. Full standard text is not reproduced here.