The Assurance Desk
What did the NCSC publish on 20 October 2026?
The National Cyber Security Centre (NCSC) published a blog post on 20 October 2026 about its research problem book. It has rewritten the book to include a new chapter called 'AI in cyber security problems'.
Much of the chapter is about agentic AI. The NCSC says agentic AI systems can plan, use tools and take actions across multiple steps.
The chapter has three broad themes. The first is using AI to strengthen cyber defence. The second is the security and assurance of AI systems. The third is using data to improve cyber resilience at scale.
A research problem book is a list of questions for researchers. The NCSC says this list is not exhaustive. It is a guide to the questions where academia, industry and government can make progress over the next 5 years.
So the chapter asks questions. It does not set rules, and it does not tell any organisation what it must do. The NCSC asks for feedback, more research questions and examples of research already under way.
What are the NCSC's AI research questions?
The NCSC says many of the assumptions behind today's cyber security practice came before this kind of AI existed. It says that ideas such as vulnerability management, incident response and risk management will need to change.
Each question in the chapter also works as a plain question you can ask a supplier or your own IT team. The table below puts them side by side.
| The NCSC's research question | A question you can ask |
|---|---|
| How can organisations redesign cyber defence workflows to get the most from agentic AI? | Did we change how the work is done, or did we add an AI agent to the old process? |
| How can AI agents reliably help defenders find, rank and reduce their attack surface? | When the tool ranks our weak points, does it know which systems matter most to the business? |
| How can AI strengthen threat detection, triage and incident response? | When the tool flags an alert, can a person see why, and decide what happens next? |
| How should autonomous or semi autonomous cyber defence systems be evaluated and assured? | Was the tool tested on systems like ours, or only in a demonstration? |
| How can individual AI and machine learning systems be secured throughout their lifecycle? | Who checks the AI system after it goes live, and how often? |
| How do we secure current and future AI ecosystems? | Which other companies' models, data and plugins does the tool depend on? |
| How can data driven methods improve cyber security analysis, posture and resilience at national scale? | This one is for national bodies, not for a single organisation. |
Why does human oversight run through the whole chapter?
The NCSC says one thread runs through all the areas: human and AI working together.
It wants to understand where autonomy adds real value, and where human judgement remains essential. It also wants to know how people can keep meaningful oversight without losing machine speed.
The NCSC also says agentic AI should not remove the need for meaningful human oversight where it matters. Research could help define when human approval is essential, and how people should review or challenge an action the AI recommends.
The NCSC has already published practical advice on this. Its blog 'Managing the cyber risk of agentic AI', from 20 August 2026, names three levels of human control. It says not every deployment needs the same degree of autonomy, and its advice should be applied in proportion.
That blog is interim advice. The NCSC says formal guidance will build on it and ultimately replace it.
Alan Woodward, writing on LinkedIn, made a similar point from the security side. He wrote that safety which lives inside the thing you are trying to secure cannot be the whole answer.
Three levels of human control over an AI agent
Human in the loop
- A person approves each action before it happens.
Human on the loop
- A person monitors the actions and can step in if needed.
Human out of the loop
- The AI acts on its own, with no human review.
An AI cannot police itself. That duty remains ours.
Agentic AI in cyber security: what is still unproven?
Some things about agentic AI in cyber security are still unproven, and the chapter says so. A system that performs well in a benchmark or a demonstration may behave differently on live systems.
The NCSC also says poor AI support could increase false confidence, miss important context or overwhelm analysts with low quality recommendations.
Its August blog adds that safety controls built into AI models give a baseline level of protection. It says they should not be treated as the whole answer.
The chapter says assurance must cover more than how well a system performs. It should cover logging, monitoring, explainability, human oversight, misuse, loss of control and misalignment, and it should continue after the system goes live.
The NCSC says it has evidence that attackers use capable AI tools, and that AI makes some attacks easier and others faster. Oliver Patel, writing on LinkedIn about AI governance, put the wider debate in proportion. He wrote that for most organisations, the doomsday scenarios are not the most relevant AI challenge they are managing.
For most organisations, the doomsday scenarios are not the most relevant AI challenge they are managing.
What does this mean if you buy AI security tools or services?
If you buy AI security tools or services, start from one point: defenders cannot use AI as freely as attackers do. The NCSC says they must work within organisational, operational and governance limits.
To benefit fully from agentic AI, the NCSC says organisations usually need to redesign their workflows, not simply add agents to them.
Peter Haigh, the NCSC's Deputy Chief Technology Officer, wrote about agentic AI in cyber defence workflows.
The NCSC says research is still needed on how permissions, approval points, logging and recovery should work for AI agents in security tools. These are fair questions for a supplier. Ask what the tool may do, who approves its actions, what it records and how a mistake is undone.
The NCSC says autonomous systems need logging, monitoring and review. Then an organisation can see what the system did, why it acted and whether it is changing.
Its August blog gives practical steps. It says AI agents should always run in a sandbox that controls what they can connect to. Where mistakes would have serious consequences, it recommends naming the people responsible for the agent's activity.
If an incident is found, it says you should always be able to 'pull the plug' and stop the agent at once.
Questions to ask before an AI tool acts in your security
- What is the tool allowed to change on its own?
- Which actions need a person to approve them?
- What does it log, and who reads the logs?
- How do we undo an action that was wrong?
- Can we stop it at once if something goes wrong?
- Does it run in a sandbox that limits what it can connect to?
- Who is named as responsible for what it does?
- Was it tested on systems like ours, or only in a demonstration?
- Who checks it after it goes live?
Safe and effective adoption of agentic AI into cyber defence workflows is a challenge facing organisations today.
What does this mean for a smaller organisation?
For a smaller organisation, the NCSC's point about incidents matters most. It says the hardest problem in many incidents is not collecting information. It is understanding what matters and deciding what to do next. Speed from AI is only valuable if it leads to better decisions.
Harry W, the NCSC's Technical Director for Incident Management, wrote about smaller organisations with limited security resources.
If your organisation only uses everyday AI tools, most of the chapter is not aimed at you. The NCSC's August blog makes a point that still applies. Not every use of AI needs the same degree of autonomy.
If you want outside help with cyber security, the NCSC runs schemes that list approved firms. The Assurance Desk recorded the NCSC's listings on 19 September 2026. Firms listed then as NCSC Cyber Advisors included Arcanum Cyber Security, Logiq Consulting and RightCue Consulting Services.
For smaller organisations with limited security resources, getting this right could make the difference between identifying a compromise early and discovering it long after the damage has been done.
What to do
- Read the NCSC's new chapter yourself. It is a set of research questions, not a rulebook.
- Decide how much autonomy each AI tool really needs, and which of its actions need a person to approve them.
- Ask suppliers how their tool was tested, and whether it was tested outside a demonstration.
- Ask what the tool logs, who reviews it after it goes live, and how you can stop it at once.
- If you have research or experience to share, the NCSC asks for feedback on the chapter.
Common questions
What is agentic AI?
The NCSC says agentic AI systems can plan, use tools and take actions across multiple steps.
Is the NCSC research problem book guidance I must follow?
No. It is a guide to research questions for academia, industry and government over the next 5 years. The NCSC says the list is not exhaustive.
Does the NCSC say AI should run cyber defence without people?
No. The NCSC says agentic AI should not remove the need for meaningful human oversight where it matters.
Does AI make incident response better because it is faster?
Not on its own. The NCSC says speed is only valuable if it leads to better decisions.
Where next
Sources
- NCSC, The NCSC research problem book, AI in cyber security problems, checked 22 October 2026
- NCSC, Managing the cyber risk of agentic AI, checked 22 October 2026
- Alan Woodward on LinkedIn, AI Safety Requires More Than Trained Models, Alan W. posted on the topic, checked 22 October 2026
- Oliver Patel on LinkedIn, Agentic AI Governance Frameworks and Guidelines, Oliver Patel, AIGP, CIPP/E, MSc posted on the topic, checked 22 October 2026
- NCSC, The NCSC research problem book, AI in cyber security problems, AI1, How can organisations re-design cyber defence workflows?, checked 22 October 2026
- NCSC, The NCSC research problem book, AI in cyber security problems, AI4, How should autonomous cyber defence systems be evaluated and assured?, checked 22 October 2026
- NCSC, The NCSC research problem book, AI in cyber security problems, AI3, How can AI strengthen threat detection, triage and incident response?, checked 22 October 2026
- NCSC, Introducing a new AI chapter in the NCSC research problem book, checked 22 October 2026
- The Assurance Desk, Arcanum Cyber Security (Arcanum Information Security Limited) profile, checked 22 October 2026
- The Assurance Desk, Logiq Consulting Ltd profile, checked 22 October 2026
- The Assurance Desk, RightCue Consulting Services Ltd profile, checked 22 October 2026
This is an independent editorial guide, not legal advice.