Skip to content

3 November 2026 | 8 min

ISO/IEC 27090: Why AI Requires Its Own Security Controls

Artificial intelligence is transforming not only business processes but also the security risks companies face. Traditional cybersecurity protects networks, user accounts, applications and data. However, AI systems introduce additional attack surfaces: training data can be manipulated, models can be stolen, inputs can be deliberately exploited and sensitive information can be extracted through seemingly legitimate requests. This is precisely where the new ISO/IEC 27090:2026 comes in. Published in October 2026, the international standard provides guidance on how organisations can understand, identify and mitigate AI-specific security threats throughout the entire lifecycle of an AI system.

ISO/IEC 27090 is designed for organisations that develop or use AI systems. The standard complements existing cybersecurity approaches by addressing risks specific to artificial intelligence. These include data poisoning, model theft or model exfiltration, prompt injection, model inversion and other attacks targeting data, models and outputs. ISO/IEC 27090 does not replace ISO 27001 or an AI management system. Instead, it extends existing security practices with guidance on AI-specific protective measures.

Why Traditional Cybersecurity Is Not Always Enough for AI

Many security frameworks focus on protecting systems, networks and access points against known attack methods. With AI systems, however, the underlying data itself can become a target. If training data is deliberately manipulated, the behaviour of a model can change even when the technical infrastructure surrounding it remains completely intact. ISO refers to this type of attack as data poisoning. Attackers attempt to introduce unwanted or malicious data into training or learning processes to influence future results.

Additional threats target the model itself. Through carefully designed queries, attackers may attempt to extract information about a model or reconstruct parts of it. ISO/IEC 27090 therefore addresses techniques such as model exfiltration, model inversion and membership inference. Prompt injection is another relevant threat. In these attacks, inputs are designed to make an AI system disregard existing instructions or perform unintended actions. These examples demonstrate that the security boundaries of an AI system cannot simply be protected through traditional firewalls, access controls or endpoint security.

Security Must Cover the Entire AI Lifecycle

One of the central principles of ISO/IEC 27090 is its focus on the entire lifecycle of an AI system. Security risks do not arise only during development. Data collected during operation may later be reused for model updates or retraining. If that data has been manipulated, vulnerabilities may only become visible in a subsequent model version. This creates additional attack surfaces, particularly for systems that continuously or regularly learn from new information.

Organisations therefore need to understand where their training and operational data comes from, how its integrity is verified and who is authorised to modify models or datasets. Development environments, model versions and updates must also be included in structured security management. This automatically makes AI security a governance issue: responsibilities, approvals, monitoring and documented controls must remain effective throughout the entire lifecycle.

Prompt Injection Is Only Part of the Problem

Prompt injection has attracted significant attention with the rise of generative AI. For example, a user may attempt to manipulate a language model into bypassing security rules or disclosing confidential information. However, ISO/IEC 27090 addresses a much broader range of threats.

These include attacks on model behaviour, manipulated training data, sensitive outputs, compromised inputs and vulnerabilities within the software and supply chain of an AI system. The standard therefore also considers threat modelling, red teaming, zero-trust principles, software engineering and supply chain security. The underlying principle is defence in depth: no single security measure can prevent every AI-related threat. Instead, multiple layers of protection must work together.

What Role Does ISO 27001 Play?

ISO/IEC 27090 is particularly relevant for companies that already operate an Information Security Management System (ISMS) based on ISO 27001. ISO explicitly describes the new standard as a complement to ISO/IEC 27001 and ISO/IEC 27002. Existing information security processes therefore remain relevant, and organisations do not need to start from scratch.

An established ISMS already provides essential foundations such as risk management, asset management, access controls, supplier management, incident management and continuous improvement. ISO/IEC 27090 helps organisations apply these structures to AI-specific threats and address additional risks that traditional information security frameworks may not fully cover.

This creates an important connection between traditional information security management and AI governance. Companies should not manage AI systems outside their existing GRC structures but integrate them into established risk and control processes.

How Does ISO 42001 Fit In?

ISO/IEC 42001 also plays an important role in this context. While ISO 27001 defines a management system for information security, ISO 42001 focuses on the management of artificial intelligence. ISO/IEC 27090 has a more specific purpose: securing AI systems and addressing the threats associated with them.

The standards therefore fulfil different functions. ISO 42001 addresses AI governance and management at an organisational level. ISO 27001 provides the general information security framework. ISO/IEC 27090 complements these structures with detailed guidance on AI-specific security risks. ISO explicitly states that ISO/IEC 27090 does not define a standalone AI management system.

For organisations, this means that instead of establishing multiple disconnected management systems, requirements should be integrated into a shared governance structure wherever possible.

AI Security Is Becoming a GRC Priority

As companies integrate AI into increasingly business-critical processes, AI security can no longer be treated purely as an IT issue. A compromised AI system may produce incorrect decisions, disclose confidential information, manipulate business processes or create regulatory risks. As a result, AI security directly affects risk management, compliance, data protection and corporate governance.

GRC leaders therefore need to know which AI systems are being used within their organisations, what purpose they serve, which data they process and what risks they introduce. Clear responsibilities for approvals, updates, security incidents and external AI providers are equally important. Only when this information is brought together in a structured way can AI security be managed effectively.

External AI Providers Create Additional Risks

Many companies do not develop their own AI models but rely on external platforms, APIs or SaaS services. However, this does not eliminate security risks. Instead, some of those risks shift to the provider.

Organisations must therefore also assess their AI suppliers. Important considerations include data usage, model updates, security measures, access permissions and incident response processes. Changes made to a model by its provider may also affect existing risk assessments.

ISO/IEC 27090 therefore addresses supply chain security as well. For GRC leaders, this means that AI governance and third-party risk management are becoming increasingly interconnected.

Security Controls Must Remain Effective Over Time

Another important aspect of the standard is the long-term effectiveness of security measures. AI systems evolve. Models are updated, datasets grow and new attack techniques emerge. A one-time security assessment is therefore not sufficient.

ISO/IEC 27090 addresses areas such as logging, monitoring, protection of development environments and the detection of malicious inputs. Existing safeguards must also be reviewed regularly, as their effectiveness may decline over time.

This reflects a fundamental principle of modern GRC systems: a control is not effective simply because it was implemented once. It must be continuously monitored and adapted to emerging risks.

Conclusion

ISO/IEC 27090 demonstrates that artificial intelligence introduces new information security requirements. Traditional cybersecurity remains essential, but it is not sufficient on its own. AI systems have additional attack surfaces, ranging from manipulated training data and model theft to prompt injection.

For companies, the challenge is therefore not to build an entirely new security management framework. Instead, existing processes based on ISO 27001, risk management and governance need to be extended to address AI-specific threats.

The central question is no longer simply: Is our IT secure?

It is also: Do we understand and control the risks of the AI systems that make decisions, process data and influence business operations within our IT environment?

FAQ

What is ISO/IEC 27090?

ISO/IEC 27090:2026 is an international standard providing guidance on identifying, assessing and mitigating security threats specific to AI systems throughout their lifecycle.

When was ISO/IEC 27090 published?

The first edition was published as an international standard on 14 October 2026.

Which risks does ISO/IEC 27090 address?

The standard covers threats such as data poisoning, model exfiltration or model theft, model inversion, prompt injection, manipulated inputs and other attacks targeting AI data, models and outputs.

Does ISO/IEC 27090 replace ISO 27001?

No. ISO/IEC 27090 complements ISO 27001 and ISO 27002 by addressing AI-specific security risks. Existing information security processes remain relevant.

Is ISO/IEC 27090 an AI management system like ISO 42001?

No. ISO/IEC 27090 provides security guidance for AI systems, while ISO/IEC 42001 defines an artificial intelligence management system.

Which companies should consider ISO/IEC 27090?

The standard is relevant to organisations of all sizes that develop or use AI systems. It is not limited to generative AI or chatbots.

Related posts