In October 2026, ISO and IEC published ISO/IEC 27090:2026. It is the first edition of a standard on security threats to artificial intelligence (AI) systems. Its full title is “Cybersecurity — Artificial intelligence (AI) — Addressing security threats and compromises to AI systems”.
If your organisation builds AI systems, or uses them, this is a useful reference. You can read it alongside the security work you already do.
What’s changing
The standard gives guidance on how to understand, detect and reduce security threats to AI systems throughout their life cycle. ISO lists it as published on 14 October 2026.
It is guidance, not a set of requirements. ISO’s guide to the standard says it “provides security guidance, not requirements for an AI management system”. So there is no certificate to gain and no deadline to meet.
The threats it covers include:
- Data poisoning: putting malicious or unwanted data into training data.
- Model poisoning: tampering with the model directly.
- Evasion: inputs designed to make a model give the wrong answer.
- Model theft: extracting a model, sometimes through queries that look legitimate.
- Membership inference and model inversion: working out what data a model was trained on.
- Prompt and output injection: instructions hidden in inputs, or outputs that carry an attack onward.
The standard also covers supply chain security, logging and monitoring, threat modelling and red teaming. It notes that models and their safeguards can weaken over time.
It is organised in three main clauses. These cover applying information security to AI, the types of compromise, and mitigations and how they interact. Annexes map attacks to life cycle stages and assets, and cover AI versions of conventional attacks.
The scope is deliberately narrow. According to ISO27001security.com, the standard concentrates on deliberate attacks on AI systems. It does not cover attackers using AI against other systems, or the use of AI for defence.
What it means for you
ISO/IEC 27090 builds on standards you may already use. ISO’s guide says it treats ISO/IEC 27001 practices as a foundation. It adds AI-specific detail to the controls guidance in ISO/IEC 27002. It complements ISO/IEC 42001, which sets requirements for an AI management system.
A sensible way to start:
- Find your AI systems. List the AI systems you build, buy or use, and who owns each one.
- Include your suppliers. Many AI systems come from providers or rely on outside models and data.
- Assess the AI-specific risks. Use the standard’s threat types as prompts in your next risk assessment.
- Check your existing controls. Ask whether they cover training data, models and prompts, not only servers and accounts.
- Plan for change. Review AI risks when models are retrained or replaced, not only once a year.
How RiskWitness helps
You can work through the standard in RiskWitness using the register you already keep.
Record your AI systems as assets. Give each one an owner, and say where it is kept or hosted. Link it to the risks, controls and suppliers that affect it.
Add the AI risks. Use the standard’s threat types as a starting list, such as data poisoning or prompt injection. Give each risk an owner, and score likelihood and impact before and after controls.
Link risks to controls. Many AI threats are reduced by controls you already have from ISO/IEC 27001:2022. Link them, record how effective each one is, and test them.
Add the standard as your own framework if you want to. On the Enterprise plan, you can import your own frameworks from a spreadsheet. ISO/IEC 27090 isn’t in the RiskWitness library, so you would import the parts you use.
Bring in your AI suppliers. Rate each provider’s criticality. Send questionnaires about how they protect models and training data, and keep their assurance evidence.
Decide, and act. Record a treatment decision for each risk: reduce, transfer, avoid or accept. Acceptances need a reason and an expiry. Outside tolerance, a named approver signs them off. Turn gaps into actions with owners and due dates.
Report it. Give the board a PDF report or slide deck on your AI risks.
Every step is kept in a history that shows who did what, when, and the old and new values. Nobody can edit or delete it. The board, an auditor, a customer or a regulator may ask how you assessed your AI systems. The answer is already written down.
Sources
- ISO/IEC 27090:2026 Cybersecurity — Artificial intelligence (AI) — Addressing security threats and compromises to AI systems, ISO, October 2026
- ISO/IEC 27090 explained, ISO, 3 September 2026
- ISO/IEC 27090, ISO27001security.com, updated 19 October 2026