Skip to content
Atesto
Create account
All articles

Standards

ISO/IEC 27090:2026 sets technical security controls for AI

ISO and IEC published ISO/IEC 27090:2026, a 56 page standard defining technical controls for AI threats like data poisoning and prompt injection.

On October 14, 2026, the International Organization for Standardization and the International Electrotechnical Commission (opens in a new tab) published ISO/IEC 27090:2026. Titled Cybersecurity, Artificial intelligence, Addressing security threats and compromises to AI systems, the 56 page document is a guidance standard. It defines technical detection and mitigation controls across the full artificial intelligence lifecycle. The publication provides a structured framework to protect systems against vulnerabilities that emerge from data collection through model retraining.

Who it applies to

The standard applies to organizations of all types and sizes that develop or use artificial intelligence systems. This scope covers public and private companies, government entities, and nonprofit organizations. It serves security teams, developers, data scientists, software engineers, risk managers, governance teams, and procurement professionals who evaluate supplier controls.

The introduction points to sectors where failures carry severe consequences, including healthcare, robotics, and the automotive industry.

What it requires

ISO/IEC 27090:2026 is a guidance standard. It provides a framework of technical controls that organizations can implement to support broader governance efforts. It serves as a complementary standard to ISO/IEC 42001 (opens in a new tab) for artificial intelligence management systems and ISO/IEC 27001 for information security management.

Developed under subcommittee ISO/IEC JTC 1/SC 27, the document defines technical controls for specific artificial intelligence threat categories. The standard organizes its guidance around the application of information security, the types of potential compromise, and the necessary mitigations.

Clause 5 directs organizations to apply established information security controls across the entire system lifecycle. This means securing the software, hardware, and network environments surrounding the model, rather than just the model itself. The text recommends an artificial intelligence governance program with an inventory of systems and their uses. Organizations need visibility over where artificial intelligence is used before they can manage exposure and accountability.

A core component of Clause 5 is the application of zero trust principles. The text advises organizations to assume breach, verify all requests, use end to end encryption, apply least privilege, and authenticate every action. This applies to training data, model access, retrieval augmented generation, and autonomous agents. A model should not be trusted simply because a request originates from an internal system or an authenticated user.

The standard bridges traditional software engineering and artificial intelligence development. It recommends practices like threat modeling, red teaming, code review, and penetration testing. It pairs these with specific measures like continuous validation, tests for staleness, and tracking training data provenance. Artificial intelligence security failures often begin in data and model development, not only at runtime.

Clause 6 organizes compromises by the asset affected and the security property compromised. The text provides a summary of these relationships and shows whether an attacker is required. It also specifies whether the case applies to generative artificial intelligence, predictive artificial intelligence, or both. The standard covers examples such as data poisoning attacks, where an attacker manipulates data to affect behavior. It details prompt injection attacks, where crafted input instruction content triggers unintended behavior. It also explains evasion attacks, membership inference, model exfiltration, model inversion, direct model poisoning, direct model theft, direct training data leaks, model input leaks, sensitive model outputs, and output containing injection attacks. This structure helps readers match specific threats to the correct control for confidentiality, integrity, or availability.

Clause 6 also covers operational considerations. These include data limitation, oversight, least model privilege, transparency, continuous validation, explainability, and unwanted bias testing. These elements highlight that many failures are operational or trust issues that manifest during normal use.

Clause 7 emphasizes that mitigations must be treated as an interconnected system. Defenses can interact with each other, weaken other controls, and lose effectiveness over time. The standard recommends continuous logging, development environment protection, malicious input detection, filtering, throttling model use, and optimizing model size. It stresses the continuity of mitigations across the lifecycle and ongoing validation when models or defenses deteriorate. Security must continue after deployment. Operational data often feeds future updates or retraining. According to commentary from Internet Security Auditors (opens in a new tab), if that data is manipulated, it can introduce vulnerabilities into later versions.

The document includes three annexes to support implementation rather than adding new requirements. Annex A maps attacks to the system lifecycle and assets, including tables and formulas for training and inference relationships. Annex B provides examples of artificial intelligence versions of conventional attacks, such as denial of service, supply chain attacks, and backdoor attacks. Annex C explains federated machine learning, detailing its workflow, benefits, and challenges.

The dates

ISO/IEC 27090:2026 was officially published on October 14, 2026. As a published guidance standard, organizations can adopt its recommendations immediately to inform their security architectures and risk assessments.

What to check now

  • Inventory all artificial intelligence systems currently in development or production across the organization.
  • Map the data, models, and suppliers associated with each system to understand the full supply chain.
  • Verify the provenance of training data and restrict modification privileges to authorized personnel only.
  • Confirm that development environments are strictly separated from production environments to prevent unauthorized changes.
  • Establish a documented process to review operational data for anomalies before using it to retrain models.
  • Integrate artificial intelligence threat modeling and red teaming into existing information security risk assessments.
  • Review current access controls to ensure zero trust principles apply to model access and retrieval augmented generation systems.

Sources

  1. International Organization for Standardization and the International Electrotechnical Commission (opens in a new tab) · standards.iteh.ai
  2. Kertos (opens in a new tab) · www.kertos.io
  3. Internet Security Auditors (opens in a new tab) · blog.isecauditors.com

This article is general information, not legal advice. For your company's situation, talk to a qualified lawyer.

This article was written with AI and checked, sentence by sentence, against the sources linked above.