NCSC AI chapter: what it asks and what is publishedNCSC AI chapter: 7 problems
An interactive chart of the NCSC's new AI chapter as read on Tuesday 20 October 2026 and subject to change. The chapter has seven problems, AI1 to AI7, and 33 strands, counted from the page headings (derived). This site groups them in three themes, which the pages do not label: using AI for defence has 17 strands, the security of AI systems 11 and data methods at national scale 5 (derived). A status says what the UK documents read give for each problem: AI5 a published code or guidance, AI6 a code in part, AI1 and AI2 interim advice only, AI3 guidance from 2019 under review, AI4 and AI7 none found in the pages read. One tab counts strands by status, one shows the pages the NCSC's blog says will need to evolve and the AI terms found in them, and a strip lists what is not stated. Hover, tap or focus any item for the words and the source.
Seven problems, each square one strand: 33 strands (derived)
The seven problems in the NCSC AI chapter as read on 20 October 2026, with the strands counted from the page headings, this site's theme and the status of what the UK documents read give.
- AI1: Redesign defence workflows for agents. 4 strands, Using AI for defence, Interim advice only.
The NCSC asks how organisations can redesign defence workflows to get the benefit of agentic AI.
NCSC problem page AI1, published 20 October 2026: “how it should interact with existing tools, and how humans can retain appropriate oversight and control”
NCSC blog on managing the cyber risk of agentic AI, published 20 August 2026: “the NCSC is sharing interim practical advice” “will build upon, and ultimately supersede, this blog” The blog says formal guidance is being developed with partners.
4 strands counted from the page headings (derived).
The nearest UK documents read are two NCSC blogs, of 20 August and 21 September 2026. Neither redesigns a defence workflow.
Not stated: a redesigned defence workflow, or a date for the formal guidance.
- AI2: Agents reduce the attack surface. 5 strands, Using AI for defence, Interim advice only.
The NCSC asks how AI agents can reliably help defenders find, rank and reduce an attack surface.
NCSC problem page AI2, published 20 October 2026: “Extensive research is needed before we can trust agents to reliably remediate large parts of the attack surface.”
NCSC blog on frontier AI, published 30 March 2026: “still only identify problems rather than fix them” The blog calls attack surface reduction the first of three top bets.
5 strands counted from the page headings (derived). 0 mentions of AI, artificial intelligence, machine learning, agentic, agent or agents in the vulnerability management pages read (derived).
The nearest UK documents read are an NCSC blog of 30 March 2026 and its vulnerability management guidance, which is written for human run processes.
Not stated: guidance on agents that fix problems, not only find them.
- AI3: Detection, triage and incident response. 4 strands, Using AI for defence, 2019 guidance, under review.
The NCSC asks how AI can strengthen detection, triage and incident response.
NCSC problem page AI3, published 20 October 2026: “poor AI support could increase false confidence, miss important context or overwhelm analysts”
NCSC guidance on intelligent security tools, published 18 April 2019: “This guidance is currently under review.” Published and reviewed on 18 April 2019, version 1.0.
4 strands counted from the page headings (derived). 18 April 2019 to 20 October 2026 is 2,742 days (derived).
The guidance the NCSC points to for buying AI based security tools is its 2019 page on intelligent security tools. The 30 March 2026 blog covers detection and response as two of its bets.
Not stated: an update date for the 2019 guidance.
- AI4: Evaluate and assure autonomous defence. 4 strands, Using AI for defence, None found in the pages read.
The NCSC asks how autonomous and semi autonomous defence systems should be evaluated and assured.
NCSC problem page AI4, published 20 October 2026: “A system that performs well in a benchmark or demonstration may behave differently in a live environment”
NCSC blog on managing the cyber risk of agentic AI, published 20 August 2026: “independently evaluated for performance” The sentence is about additional judge models and agents, not about a defence system as a whole.
4 strands counted from the page headings (derived). AI4 is the only one of the 7 problem pages with no closing statement (derived).
The nearest are an NCSC blog that rates the risk of an action, not of a system, and a line in the August blog about judge models.
Not stated: a test, a baseline or a pass mark for a defence system.
- AI5: Secure AI systems across their lifecycle. 5 strands, Security of AI systems, Code or guidance published.
The NCSC asks how individual AI and machine learning systems can be secured across their lifecycle.
NCSC problem page AI5, published 20 October 2026: “A bespoke security assessment for every AI deployment is unlikely to scale.”
DSIT Code of Practice for the Cyber Security of AI, published 31 January 2025: “voluntary Code of Practice” The Code was published by DSIT, the Department for Science, Innovation and Technology.
5 strands counted from the page headings (derived). 13 principles in the Code, and 0 uses of the word agent in the Code and in the ETSI standard (derived).
Three UK documents read cover the subject: a government Code of Practice, the NCSC's guidelines for secure AI system development and an ETSI standard. They set principles for the systems. None is a method for assuring one deployment (this site's reading).
Not stated: a method to assure a single AI deployment.
- AI6: Secure AI ecosystems and supply chains. 6 strands, Security of AI systems, Code, in part.
The NCSC asks how current and future AI ecosystems can be secured.
NCSC problem page AI6, published 20 October 2026: “We still lack dependable ways to establish provenance and trust across AI supply chains”
DSIT Code of Practice for the Cyber Security of AI, published 31 January 2025: “Secure your supply chain” The title of principle 7.
6 strands counted from the page headings (derived).
Parts of the Code of Practice cover it: principle 7 on the supply chain, principle 11 on updates and principle 12 on monitoring. The joint guidance on careful adoption of agentic AI services is named by the NCSC and was not read.
Not stated: anything on multi agent ecosystems in a document read.
- AI7: Data methods at national scale. 5 strands, Data methods at national scale, None found in the pages read.
The NCSC asks how data driven methods can improve cyber security analysis, posture and resilience at national scale.
NCSC problem page AI7, published 20 October 2026: “Data may be incomplete, inconsistent, outdated, duplicated or difficult to match across sources.”
NCSC Cyber Shield blog, published 7 July 2026: “analysis of aggregated data to understand national level exposure” The blog describes it as an intended function of Cyber Shield.
5 strands counted from the page headings (derived).
The nearest document read is the Cyber Shield blog of 7 July 2026, which lists analysis of aggregated data as a planned function. It states no method.
Not stated: a method, a data model, a date or funding.
Three themes, this site's grouping
- Using AI for defence: 17 strands.
The blog names three themes, and neither it nor the chapter introduction says which problem sits under which.
NCSC blog introducing the AI chapter, published 20 October 2026: “identifies seven areas spanning three broad themes” The blog's own count of areas and themes.
4 of the 7 problems carry 17 of the 33 strands, 51.5 per cent (derived). If AI4 is read as assurance of AI systems, the split is 13, 15 and 5 (derived).
The grouping is this site's, made by order and wording. AI4 is about assuring defence systems that are themselves AI, so it can be read under either of the first two themes.
Not stated: which problem sits under which theme.
- Security of AI systems: 11 strands.
AI5 is about securing individual AI and machine learning systems and AI6 about securing AI ecosystems.
NCSC blog introducing the AI chapter, published 20 October 2026: “identifies seven areas spanning three broad themes” The blog names the security and assurance of AI systems as one of the three.
2 of the 7 problems carry 11 of the 33 strands, 33.3 per cent (derived). Strands about using AI, AI1 to AI4 and AI7, are 22 of the 33 (derived).
The grouping is this site's, made by order and wording. The blog names the security and assurance of AI systems as one of its three themes.
Not stated: which problem sits under which theme.
- Data methods at national scale: 5 strands.
AI7 asks how data driven methods can improve cyber security analysis, posture and resilience at national scale.
NCSC blog introducing the AI chapter, published 20 October 2026: “identifies seven areas spanning three broad themes” The blog's third theme is the use of data driven methods to improve cyber resilience at scale.
1 of the 7 problems carries 5 of the 33 strands, 15.2 per cent (derived).
The grouping is this site's, made by order and wording. The blog's third theme is the use of data driven methods to improve cyber resilience at scale.
Not stated: which problem sits under which theme.
Strands under each status, to scale (derived)
- Code or guidance published: AI5, 1 problem, 5 strands.
A UK government Code of Practice, the NCSC's guidelines for secure AI system development and an ETSI standard cover the subject of AI5. The Code is voluntary.
DSIT Code of Practice for the Cyber Security of AI, published 31 January 2025: “voluntary Code of Practice”
AI5 carries 5 strands (derived).
The two problems with a code behind them, AI5 and AI6, carry 11 of the 33 strands (derived).
Not stated: a method to assure one AI deployment.
- Code, in part: AI6, 1 problem, 6 strands.
Principles 7, 11 and 12 of the Code of Practice cover parts of AI6: the supply chain, updates and monitoring. Nothing read covers multi agent ecosystems.
DSIT Code of Practice for the Cyber Security of AI, published 31 January 2025: “Secure your supply chain” The title of principle 7.
AI6 carries 6 strands (derived).
The two problems with a code behind them, AI5 and AI6, carry 11 of the 33 strands (derived).
Not stated: anything on multi agent ecosystems in a document read.
- Interim advice only: AI1 and AI2, 2 problems, 9 strands.
The nearest UK documents read for AI1 and AI2 are NCSC blogs. The NCSC calls its August blog interim and says formal guidance will supersede it.
NCSC blog on managing the cyber risk of agentic AI, published 20 August 2026: “Until formal guidance is published, the NCSC is sharing interim practical advice”
AI1 and AI2 carry 4 plus 5, which is 9 strands (derived).
Not stated: a date for the formal guidance.
- 2019 guidance, under review: AI3, 1 problem, 4 strands.
The guidance the NCSC points to for buying AI based security tools is its 2019 page on intelligent security tools, which carries a note that it is under review.
NCSC guidance on intelligent security tools, published 18 April 2019: “This guidance is currently under review.”
AI3 carries 4 strands (derived). 18 April 2019 to 20 October 2026 is 2,742 days (derived).
Not stated: an update date for the 2019 guidance.
- None found in the pages read: AI4 and AI7, 2 problems, 9 strands.
None found means none in the pages read, not that none exists. For AI4 no document read sets out how to evaluate or assure an autonomous defence system. For AI7 the nearest is an intended capability on the Cyber Shield blog.
AI4 and AI7 carry 4 plus 5, which is 9 strands (derived). The five problems about using AI, AI1 to AI4 and AI7, carry 22 of the 33 strands, and none has current formal guidance written for it (derived).
Not stated: a test or pass mark for a defence system, and a method for AI7.
The pages the NCSC's blog says will need to evolve, and the AI terms found in them
- Vulnerability management: update by default, Reviewed 1 May 2026: 0 AI terms.
The blog names vulnerability management as a concept that will need to evolve. The page gives best practice timescales of 5 days for internet facing services, 7 for operating systems and applications and 14 for internal or air gapped systems. When exploitation is rife it says to go faster.
NCSC blog introducing the AI chapter, published 20 October 2026: “will need to evolve” The blog names vulnerability management, incident response and risk management.
NCSC page on updating by default, published 12 February 2024: “the timelines above are too long” “10% of the estate updated per day” The first is the page's own words for when exploitation is rife. The second is about a phased rollout.
NCSC Cyber Shield blog, published 7 July 2026: “activities that once took weeks can now take minutes”
5 days is 120 hours (derived). 0 mentions of AI, artificial intelligence, machine learning, agentic, agent or agents in the vulnerability management pages read (derived).
This chart does not say the 5 day figure is wrong. The page presents it as best practice for business as usual updates.
Not stated: a timescale for work done at the speed of an agent.
- Incident management: response processes, Reviewed 19 September 2019: 0 AI terms.
The blog names incident response as a concept that will need to evolve. The page plans for escalation to people, contacts including out of hours, and playbooks that cover the first few hours.
NCSC blog introducing the AI chapter, published 20 October 2026: “will need to evolve” The blog names vulnerability management, incident response and risk management.
NCSC page on incident response processes, reviewed 19 September 2019: “a high or critical severity incident is likely to always need to go up to CIO or board level” “cover at least the first few hours of each incident”
0 mentions of AI, artificial intelligence, machine learning, agentic, agent or agents in the incident management pages read (derived).
Not stated: what happens when the first few hours pass in minutes.
- Risk management collection, Reviewed 23 June 2023: 0 AI terms.
The blog names risk management as a concept that will need to evolve. The collection is guidance for risk practitioners and says it is not prescriptive.
NCSC blog introducing the AI chapter, published 20 October 2026: “will need to evolve” The blog names vulnerability management, incident response and risk management.
NCSC risk management collection, published 23 June 2023: “not prescriptive”
0 mentions of AI, artificial intelligence, machine learning, agentic, agent or agents in the risk management page read (derived).
Not stated: how to rate the risk of an agent.
- Cyber Essentials requirements v3.3, April 2026: 0 AI terms.
The scheme sets updates for critical or high risk fixes within 14 days. Like the update by default page, its timescale is counted in days.
Cyber Essentials requirements for IT infrastructure, version 3.3, April 2026: “14 days is considered a reasonable period” “a shorter period may not be practical”
0 mentions of AI, artificial intelligence, machine learning, agentic, agent or agents in the Cyber Essentials requirements (derived).
Not stated: any mention of AI.
- Intelligent security tools, Reviewed 18 April 2019: About AI, not counted.
The subject of this guidance is AI, so its AI terms are not counted. It is the guidance the NCSC points to for AI based security tools, and it carries a note that it is under review.
NCSC guidance on intelligent security tools, published 18 April 2019: “This guidance is currently under review.” Published and reviewed on 18 April 2019, version 1.0.
18 April 2019 to 20 October 2026 is 2,742 days (derived).
Not stated: a date for the update.
Four sets of pages read: 12,169 words, 0 AI terms (derived). The AI terms searched for were AI, artificial intelligence, machine learning, agentic, agent or agents.
Not stated in the sources read
Any duty on a UK organisation
The chapter calls its list a guide to research questions. The word must appears twice in the eight pages, neither time as a duty on an organisation, and shall appears 0 times.
NCSC chapter introduction, published 20 October 2026: “a guide to the types of research questions”
must appears 2 times and shall 0 times in the eight pages (derived).
Not stated: any duty on a UK organisation, or any link to a regulator, a law or a procurement rule.
An end year, or a date for an answer
The chapter says the next 5 years. The book introduction says the next decade and five to ten years.
NCSC chapter introduction, published 20 October 2026: “the next 5 years”
NCSC problem book introduction, read 20 October 2026: “the next decade” “five to ten years”
5 years from 20 October 2026 would end in 2031, a year no page prints (derived).
Not stated: an end year, a review date, or a date for an answer to any of the seven.
Funding, a budget, a grant or a call
The blog welcomes feedback, suggestions and examples of research and points to a form. The words fund and grant appear 0 times on the eight pages.
NCSC blog introducing the AI chapter, published 20 October 2026: “We welcome feedback on the challenges identified in the problem book”
fund and grant appear 0 times in the eight pages (derived).
Not stated: funding, a budget, a grant or a call.
A measure or pass mark for an answer
AI2 wants metrics that are transparent and meaningful. The words threshold and deadline appear 0 times on the eight pages.
NCSC problem page AI2, published 20 October 2026: “transparent and meaningful” The page says the metrics should support better decisions.
threshold and deadline appear 0 times in the eight pages (derived).
Not stated: a measure, a threshold or a pass mark for reliably, meaningful or secure enough.
Any UK code or standard named
The eight pages build on two NCSC blogs and a form. The seven problem pages carry 0 outbound links.
NCSC chapter introduction, published 20 October 2026: “we have evidence that capable AI tools are being used by threat actors” The evidence is not cited on the page.
The Code of Practice, the Guidelines for secure AI system development and ETSI are each named 0 times in the eight pages (derived). 0 outbound links in the 7 problem pages (derived).
Not stated: the Code of Practice, the Guidelines or the ETSI standard, named anywhere.
What happens to a submission
The form asks for a name, an email address, a title, a summary, the area and any previous relationship with the NCSC. Nothing was typed into it while the page was read.
NCSC problem book feedback form, read 20 October 2026: “250 words or fewer” The form's limit for the summary.
Not stated: a reply, a publication, a review time, or how a result would reach guidance.
Derived means this site's arithmetic. Theme and status are this site's reading of the NCSC pages read on 20 October 2026.
Derived means this site's arithmetic.