ISO/IEC 27090:2026 Explained: AI Security | GRC Link

What the new AI cybersecurity standard means for organisations developing and using artificial intelligence.

AI is changing how organisations make decisions, serve customers and manage operations. But are existing cybersecurity measures equipped to protect the systems behind those decisions?

Artificial intelligence (AI) introduces security risks that extend beyond conventional software and IT infrastructure. An attacker may target the data used to train a model, manipulate the instructions it receives or attempt to extract sensitive information from its outputs.

Published on 14 October 2026, ISO/IEC 27090:2026 provides guidance on understanding, detecting and mitigating security threats and compromises specific to AI systems.

For organisations developing or using AI, the question is no longer simply whether their systems are secure. It is whether their approach to information security accounts for how AI systems learn, behave and change over time.

1. What Is ISO/IEC 27090:2026, and Why Does It Matter?

The standard, formally titled Cybersecurity — Artificial intelligence (AI) — Addressing security threats and compromises to AI systems, focuses on threats associated with AI technologies throughout their lifecycle.

Traditional information security practices remain essential. Access management, secure infrastructure, monitoring and protection of information assets do not become less important when AI is introduced.

However, AI systems depend on training data, models and changing inputs. This introduces additional ways for malicious activity to influence their behaviour or compromise the information they process.

ISO/IEC 27090 complements established cybersecurity practices rather than replacing them. Its purpose is to help organisations recognise AI-specific vulnerabilities, understand their potential consequences and consider appropriate detection and mitigation measures.

This is relevant to businesses using AI in customer service, finance, manufacturing, healthcare, decision support and other operational environments. Where AI influences critical decisions, compromised behaviour can affect not only information security but also organisational reliability and, in some cases, human safety.

2. What Makes AI Security Different from Traditional Cybersecurity?

Conventional cybersecurity often concentrates on protecting networks, applications, devices and access to information. AI security must also consider the integrity of training data, the confidentiality of models and the reliability of model outputs.

Consider a system trained to recognise operational anomalies. If the data used to train or update that system is deliberately manipulated, its recommendations may become unreliable without an obvious compromise to the underlying IT infrastructure.

That is why AI security requires attention to data, models, inputs, outputs and the systems surrounding them.

Six AI security threats organisations should understand

1. Data poisoning: Malicious or misleading information is introduced into training or retraining data to influence how an AI system behaves. This may affect the accuracy and reliability of its outputs.

2. Model theft and extraction: Attackers attempt to copy, reconstruct or obtain access to proprietary AI models, potentially compromising valuable intellectual property.

3. Prompt injection: Crafted inputs attempt to redirect an AI system away from its intended instructions or controls. In some contexts, this could lead to inappropriate outputs or information exposure.

4. Membership inference attacks: An attacker attempts to determine whether particular information was included in a model’s training data, creating potential confidentiality and privacy concerns.

5. Model inversion attacks: An attacker uses model outputs or other available information to infer or reconstruct sensitive characteristics of the data on which the model was trained.

6. Evasion attacks: Carefully designed inputs cause an AI system to produce incorrect classifications or decisions even though the underlying system continues to operate.

These threats can affect confidentiality, data integrity, decision accuracy, intellectual property and organisational trust. Where AI supports safety-critical processes, the consequences may be more serious.

3. How Does ISO/IEC 27090 Fit with ISO/IEC 27001, 27002 and 42001?

Organisations do not need to treat AI security as a completely separate discipline. ISO/IEC 27090 is intended to complement existing information security and AI governance approaches.

The related standards serve different but connected purposes:

StandardPrimary focusHow it supports AI security
ISO/IEC 27001Information security management systemsEstablishes a structured approach to managing information security risks.
ISO/IEC 27002Information security controlsProvides guidance on controls that protect information assets.
ISO/IEC 42001AI management systemsSupports governance, accountability and the responsible management of AI.
ISO/IEC 27090AI-specific cybersecurity threatsAddresses threats to AI data, models and workflows, together with relevant detection and mitigation considerations.

The value lies in how these standards work together.

ISO/IEC 27001 and 27002 support established information security practices. ISO/IEC 42001 addresses organisational governance of AI. ISO/IEC 27090 brings greater focus to the specific ways AI systems may be attacked or compromised.

For organisations already aligned with ISO/IEC 27001 or considering ISO/IEC 42001, this provides a practical basis for reviewing whether existing policies, risk assessments and controls sufficiently account for AI-related threats.

4. Five Practical Steps to Strengthen AI Security

Understanding emerging threats is important. Translating that understanding into organisational action is where security and governance become effective.

Step 1: Extend risk assessments to AI-specific threats

Review the AI systems your organisation develops, purchases or uses, including their models, training data, integrations and associated workflows.

Consider threats such as data poisoning, model extraction, prompt injection and unintended disclosure of sensitive information. Identify who is responsible for assessing these risks and how they will be monitored.

Step 2: Strengthen training data integrity

Reliable AI outputs depend on the quality and integrity of the underlying data.

Review where training and retraining data comes from, how it is validated and whether unauthorised changes could go undetected. Data poisoning can undermine a model’s behaviour before the resulting errors become obvious.

Step 3: Protect AI models and intellectual property

AI models can represent substantial business investment and may contain information that organisations need to protect.

Assess access controls, permitted model use, monitoring arrangements and safeguards against unauthorised extraction, copying or replication.

Step 4: Embed security throughout the AI lifecycle

AI security should not begin and end at deployment.

Protection needs to be considered during development, testing, implementation, operation, monitoring, model updates, retraining and eventual retirement. A system that was secure at launch may develop new exposures as data, integrations or usage patterns change.

Step 5: Integrate AI security into existing governance

Managing AI risks in isolation can create fragmented responsibilities and inconsistent controls.

Where possible, integrate AI-specific security considerations into existing information security, risk management, privacy and governance structures. This gives leadership clearer accountability and helps ensure that AI-related risks are considered alongside wider organisational priorities.

5. What Should Business Leaders Be Asking?

For many organisations, AI adoption is advancing faster than the formal processes responsible for governing its use.

Whether an organisation is building its own AI solutions or adopting third-party tools, leadership should be able to answer several practical questions:

  • Do we know which AI systems are being used, what information they access and who is accountable for them?

  • Have our information security risk assessments considered threats to AI models, data and outputs?

  • Do we understand how model updates, retraining and integrations could introduce new vulnerabilities?

  • Can we identify suspicious AI-related behaviour and respond effectively if something goes wrong?

  • Are our existing governance and security arrangements keeping pace with how AI is actually used?

These questions should form part of a broader organisational discussion, not simply an IT checklist.

6. The GRC Link Perspective: Security, Governance and Responsible Innovation

AI creates opportunities to improve efficiency, analyse information and support better decisions. But innovation without appropriate governance introduces avoidable uncertainty and risk.

At GRC Link, we believe technology should strengthen organisational performance while preserving the accountability, information security and professional judgement on which effective management systems depend.

ISO/IEC 27090 reinforces an important principle: AI security is not separate from governance. It is part of responsible AI adoption.

Organisations that connect cybersecurity, risk management, AI governance and operational processes are better positioned to protect their data, intellectual property, customers and reputation.

This builds on a wider challenge facing businesses today: technology may be advancing rapidly, but organisational controls, processes and user awareness must evolve alongside it.

For further context, read our discussion on who is governing AI use within organisations.

7. The Future of AI Security Starts with the Systems in Place Today

The publication of ISO/IEC 27090:2026 is an important recognition that AI systems require security considerations beyond those addressed by traditional controls alone.

It offers guidance for organisations developing or using AI to identify vulnerabilities, assess their potential impact and strengthen protection across the AI lifecycle.

Securing AI is no longer a future concern. It is a current business responsibility.

The organisations best prepared for this shift will be those that treat AI security as part of a connected approach to information security, governance, risk management and continuous improvement.

Is your organisation’s governance keeping pace with AI adoption?

Speak to GRC Link about strengthening the management systems, risk processes and information security practices that support responsible AI adoption.

Explore more insights and get in touch: grclink.online


Leave a Reply

Your email address will not be published. Required fields are marked *