Cybernion
← Insights

ISO 27090

ISO 27090 Threats and Mitigations

26 Oct 2026 · Updated 26 Oct 2026

ISO/IEC 27090:2026 sets out the security threats specific to AI systems and how to treat them. Its Clause 6 describes compromise types such as data poisoning, evasion, model theft and prompt injection, organised by the asset affected. Its Clause 7 covers mitigations, treated as an interacting system that can weaken over time and needs ongoing validation.

How does ISO 27090 describe AI threats?

It groups them by the asset affected and the security property at risk. Rather than a flat list, ISO 27090 organises AI compromises around what is being attacked, the training data, the model, the inputs or the outputs, and whether confidentiality, integrity or availability is at stake. It also notes whether a given case needs an active attacker and whether it applies to generative AI, predictive AI or both. This structure helps a team match a threat to the right control. The point the standard keeps making is that AI security failures often begin in data and model development, not only at runtime, so the threats span the whole life cycle. For the standard itself, see what is ISO 27090.

What are the main AI compromise types?

The standard describes a set of AI specific attacks. The headline categories, with where each lands:

Compromise type What it targets
Data poisoning Training data, to change model behaviour
Evasion attack Inputs at inference, to force a wrong result
Membership inference Training data confidentiality, testing if a record was used
Model inversion Reconstructing training data from the model
Model theft or exfiltration Copying or lifting the model itself
Direct model poisoning Tampering with the model directly
Training data leak Exposure of the data behind the model
Sensitive model output The model revealing data it should not
Prompt injection Crafted instructions that hijack model behaviour
Output contains injection Malicious content passed through model output

The annexes extend this with AI specific versions of conventional attacks, including denial of service, supply chain attacks and backdoors, plus a treatment of federated machine learning. Many of these overlap with the OWASP Top 10 for LLMs and the adversarial techniques catalogued in MITRE ATLAS.

What mitigations does ISO 27090 recommend?

It recommends security practice across the life cycle, not a one off fix. On the practice side, Clause 5 points to established software engineering, testing, code review, threat modelling and penetration testing, plus AI specific measures: training data provenance and lineage, model traceability, continuous validation, and tests for staleness and concept drift. It presses zero trust principles onto training data, model access, retrieval augmented generation and AI agents, so a request is verified rather than trusted because it comes from inside. Clause 7 then treats mitigations as a system: it recommends logging and monitoring, protecting the development environment, detecting and filtering malicious input, throttling model use, and optimising model size. Crucially, it warns that mitigations can interact, can weaken each other, and can lose effectiveness over time, so they need ongoing validation. The practical lessons line up with securing machine learning models and securing RAG and AI agents.

How should a team use this?

As a threat library to assess AI systems and suppliers against. A workable approach: inventory where AI is used; run the relevant compromise types above through your risk assessment; map each to a mitigation and an owner; add AI threat modelling and red teaming to assurance; and review the mitigations periodically, because the standard is clear they degrade. For organisations running a management system, these feed the risk work in ISO 42001 or an ISO 27001 ISMS. Cybernion helps teams test and harden AI systems through AI governance and security readiness and penetration testing.

Frequently asked questions

How does ISO 27090 organise AI threats?

By the asset affected and the security property compromised. It then notes whether an attacker is required and whether the threat applies to generative AI, predictive AI or both.

Does ISO 27090 give specific controls?

It gives mitigations and practices, not an auditable control list. It treats mitigations as an interacting system that can degrade over time and needs ongoing validation.

Are these threats unique to AI?

Some are AI specific, like data poisoning and model inversion. Others are AI specific versions of conventional attacks, such as supply chain and denial of service, covered in the annexes.


Written by Gaurav Vikash, an ASD endorsed IRAP assessor and senior cyber security leader with 18 years of experience across Australia, the UK and Asia, including CISO and senior security leadership roles. He holds CISSP, CISA, CISM and CRISC and is an ISO 27001 Lead Auditor and ISO 42001 Lead Implementer, and speaks regularly at industry conferences.

If you want to test your AI systems against these threats, talk to us.

Sources:

  1. ISO/IEC 27090:2026, Cybersecurity, Artificial intelligence, Addressing security threats and compromises to AI systems, October 2026
  2. ISO/IEC 27090:2026 catalogue entry, iTeh Standards, 2026
  3. MITRE ATLAS, Adversarial Threat Landscape for AI Systems, 2026

Last updated: 26 October, 2026

Talk to us about your engagement

A scoped proposal within one business day.