Detailed guidance issued on protecting AI from security threats

Developed by ISO/IEC and published in the UK by the British Standards Institution (BSI), the UK’s national standards body, the guidance recognises that AI is now becoming embedded into business operations. The BSI’s study highlights the fact that agentic AI use is already widespread, with nine in every ten companies exploring it in some form.

While established cyber security controls remain relevant, the clear message is that these controls are highly likely to need adapting or supplementing in order to address new risks.

While many organisations will be using AI to enhance their security, evidence suggests cyber criminals are exploiting vulnerabilities in AI systems and attacking them, while also using AI to facilitate conventional cyber crime.

Criminal tactics can include the manipulation of training data and inputs, theft or reconstruction of models and unintended behaviour triggered through prompts or outputs. The UK’s National Cyber Security Centre has predicted that AI will lead to an increase in the frequency and intensity of cyber threats.

Attacks and vulnerabilities

Cyber Security – Artificial Intelligence (AI) – Addressing Security Threats and Compromises to AI Systems (BS ISO/IEC 27090:2026) explains how attacks and vulnerabilities can affect AI data, models, outputs and services throughout the AI system lifecycle, describes their potential consequences and sets out approaches to detecting and mitigating the risks.

The document covers threats including data poisoning, prompt injection, model theft and the leakage of training data or model inputs. Importantly, the guidance applies to organisations of all types and sizes, including public and private sector companies, Government bodies and not-for-profit organisations.

The BSI’s detailed study found that security is the foremost concern around agentic AI, with more than half of business leaders (53%) stating this. AI cyber security was also the most common answer when asked which AI capability would become more important in the coming year, as cited by 48% of respondents to the survey.

The research, which explored the cost of AI failures, also finds that 75% of organisations have experienced an AI failure in the past year, with 30% of those affected confirming that such episodes had realised a security breach.

With 78% of businesses expecting their AI investment to increase, and deployment moving steadily from pilots to scale, exposure to security risk is only likely to grow.

BS ISO/IEC 27090 maps threats to the assets they affect and the lifecycle stages where they arise – from design and training through to deployment and operation – and combines conventional cyber security controls with AI-specific measures.

Serious security risks

Laura May Bishop, sector lead for AI and cyber security at the BSI, stated: “For all the opportunities presented by AI, there’s no doubt it introduces serious security risks that conventional cyber security measures alone were not designed to address. Cyber criminals can be relied upon to ruthlessly exploit AI so it’s very much the case that business leaders desperately need to stay one step ahead. We are already seeing instances of AI failure leading to security breaches, while it’s likely the true scale of the challenge is even greater.”

Further, Laura May Bishop observed: “BS ISO/IEC 27090 gives organisations around the world a structured way in which to understand these threats across the whole life of an AI system and decide on proportionate safeguards. Securing AI is not only a job for developers. It needs security, risk, governance and procurement teams working together.”

BS ISO/IEC 27090 is designed to work alongside frameworks that organisations may already use. For example, it sets out how to identify AI-specific risks and safeguards that can feed into information security management under ISO/IEC 27001 and ISO/IEC 27002 and AI governance under ISO/IEC 42001, rather than creating a separate management system.

The new guidance also complements ETSI EN 304 223, which sets baseline cyber security requirements for AI models and systems.

While the BSI’s data shows progress being made, AI tools are regularly deployed by businesses without appropriate governance programmes in place to protect against security issues. For example, only 38% of business leaders said their organisation has established an AI governance programme.

Organisational issue

The guidance treats AI security as an organisational issue rather than solely a technical one for developers, thereby affording security, AI engineering, risk, governance, procurement and operational teams a common language for discussing and managing AI security risk.

By building a better understanding of the security risks posed to AI systems, organisations can raise awareness across their teams, strengthen their defences and develop a stronger security culture.

*For more information visit BS ISO/IEC 27090:2026 | 31 October 2026 | BSI Knowledge