AI moved from pilot projects to everyday work faster than most security programmes could follow. Staff paste documents into chat assistants, developers connect language models to customer portals, and vendors add AI features to products you already pay for. The National Cybersecurity Authority has responded with the AI Cybersecurity Guidelines, which set cybersecurity expectations for how organizations in the Kingdom use and build AI.

The public consultation closed in August 2026 and the text is being finalized. The direction is already clear, so this is a good time to prepare instead of rushing when the final version comes out.

What the guidelines cover #

The guidelines deal with the cybersecurity of AI across its life: the AI tools an organization uses, the AI systems it develops and the AI services it buys. The scope explicitly includes generative AI, which produces text, images and code, and agentic AI, where a system can plan and act on its own, for example by sending emails, querying databases or calling other systems.

They apply to public and private entities that use AI or plan to. In practice that means almost every organization, because AI now arrives inside ordinary software as well as through dedicated projects.

The four domains #

The draft has four domains that will look familiar to anyone who knows the Essential Cybersecurity Controls. Governance covers who owns AI risk, which policies apply, how AI systems are approved before use and how leadership stays informed. Defence covers protecting the data, models, prompts and integrations behind AI systems, and controlling what AI agents are allowed to do. Resilience covers detecting misuse or failure, responding to AI-related incidents and recovering safely. Third-party cybersecurity covers the AI vendors, model providers and platforms you rely on, and the data you share with them.

If you already run a governance programme for the ECC or the NCNICC, AI security can be added to it rather than run as a separate project.

The risks behind the guidelines #

AI brings risks that traditional controls were not designed for. The common ones:

  • Confidential documents, customer records or source code pasted into public AI tools can leave your control.
  • Prompt injection: text hidden in an email, web page or document can instruct an AI assistant to ignore its rules, reveal data or take an action nobody intended.
  • An AI agent connected to email, files or business systems can do real damage if it has more access than its task needs.
  • AI can give confident answers that are wrong, which becomes a risk when its output feeds decisions, code or customer messages without review.
  • Models, plug-ins and AI platforms are third-party software with their own vulnerabilities and data practices.
  • Teams adopt AI tools without approval ("shadow AI"), so nobody knows what data goes where.

How to prepare now #

You do not need the final text to make real progress. These steps hold whatever the final wording turns out to be.

1. Build an AI inventory #

List the AI tools staff use, the AI features inside your existing software, and the AI systems you build or buy. For each, record what data it touches and who owns it. Most organizations find more AI in use than they expected.

2. Set an acceptable-use policy #

Tell staff which AI tools are approved, which data must never be entered, and when AI output needs a human review. Keep it short and clear. Our guide to using ChatGPT safely at work explains what to include.

3. Classify the data AI can reach #

Link the AI inventory to your data classification. Public information can flow freely; confidential, personal and regulated data needs firm limits. Personal data also brings in the Personal Data Protection Law.

4. Limit what agents can do #

For every AI system that can take actions, apply least privilege: read-only access where possible, approval steps for sensitive actions, and logs of what the agent did and why.

5. Test before you trust #

Before an AI feature goes live, test it the way an attacker would, including prompt injection and attempts to extract data. The OWASP Top 10 for LLM Applications is a useful checklist.

6. Bring AI vendors into your supplier process #

Ask AI providers where your data is stored and processed, whether it is used to train their models, how long it is kept and how they handle incidents. Put the answers in the contract.

7. Add AI to incident response #

Decide what counts as an AI incident, who handles it, and how you would switch off an AI feature quickly if it misbehaves.

Where ISO/IEC 42001 fits #

ISO/IEC 42001 is the international standard for an AI management system. It covers how an organization governs AI responsibly, including risk, impact and life-cycle management. The NCA guidelines focus on cybersecurity, while ISO/IEC 42001 covers the wider management system. Organizations that want an internationally recognized certificate for AI governance often use both, collecting evidence once and mapping it to each.

A realistic first 60 days #

  • Days 1 to 20: appoint an owner, build the AI inventory and publish an interim acceptable-use notice.
  • Days 21 to 40: classify the data each AI system touches, review agent permissions and question your main AI vendors.
  • Days 41 to 60: run a readiness assessment against the four domains, test your highest-risk AI feature and agree a roadmap with leadership.

Start with the inventory, since everything else depends on knowing what AI you actually run. Learn more about our AI cybersecurity readiness service and ISO/IEC 42001.

Frequently asked #

Are the NCA AI Cybersecurity Guidelines final?

The public consultation closed in August 2026 and the text is being finalized. Organizations can prepare now against the published draft and adjust when the final version is issued.

What are the four domains of the guidelines?

Cybersecurity governance, cybersecurity defence, cybersecurity resilience, and third-party cybersecurity.

Do the guidelines apply if we only use tools like ChatGPT or Copilot?

Staff use of AI tools is within the scope of AI governance, and it is one of the most common sources of data leakage. An inventory and an acceptable-use policy are the right first steps.

How do the guidelines relate to ISO/IEC 42001?

The NCA guidelines focus on AI cybersecurity, while ISO/IEC 42001 defines a certifiable AI management system. Many organizations use both and map one set of evidence to each.

Official Sources and References #

The documents behind this page, linked at the publisher. Where a Saudi regulation is published in both languages, the Arabic text is the binding version.