Budget 2026 promises a new unit to police 'rogue' AI systems. The one word that would tell you what counts as rogue is never defined.
Estimated reading time: 6 minutes
In short. Budget 2026, published by HM Treasury on 28 October 2026, sets aside £40 million "to establish a National Agentic Security Unit to investigate and respond to rogue agentic incidents" [1]. Nowhere in the document's 166 pages is a "rogue agentic incident" defined, nor is the new unit given a stated legal power, a named oversight body, or a date by which it must exist. The same section claims the government "will therefore invest around £500 million to build the secure foundations for an AI-enabled economy" [2], but its own itemised list of what that money buys adds up to £314 million. The £176 million gap closes only in the Treasury's separate scorecard, which bundles the security money with spending the Budget's own later pages describe as a different purpose: accelerating AI adoption [3]. Two documented mistakes: four stars.
Budget 2026 has already produced two other teardowns on this site: a gas standing charge pledge that slips a year beyond the Chancellor's own speech, and a rough sleeping "permanent" claim that the Budget's own scorecard funds only to 2029 [1]. Its AI security section, four pages on cyber defence and AI risk tucked into the "Building the UK's AI future" chapter, has had no teardown of its own. It is where the government announces a brand new operational body with an investigative remit, and where its own arithmetic does not hold up against its own appendix.
The claims, tested
| The Budget's own words | What the Budget's own other pages say | Verdict |
|---|---|---|
| Page 6, the Budget's headline summary: "It will invest £154 million to strengthen cyber defences and £40 million to establish a National Agentic Security Unit to investigate and respond to rogue agentic incidents" [1]. Paragraph 6.106 adds the unit will "build world-first capabilities to detect and respond to misaligned agentic activity" [2] | "Rogue agentic incident" and "misaligned agentic activity" appear nowhere else in the document and are not defined. No power to investigate anyone is stated, no named accountable body or minister is given beyond "the government", and no date is set for the unit to exist. The same paragraph, 4.131, announces two other new bodies with overlapping jobs in the same breath: a £70 million NCSC centre "to understand and reduce national cyber exposure" and a £60 million "Cyber Shield" to "develop agentic cyber defence capabilities", with no stated division of labour between the three [2] | A minister can claim credit for "investigating" and "responding" to a category of incident the Budget never defines, through a unit whose boundary with two neighbouring new bodies created in the same paragraph is never explained |
| Paragraph 4.129: "The government will therefore invest around £500 million to build the secure foundations for an AI-enabled economy" [2] | The paragraphs that immediately follow, 4.130 and 4.131, which 4.129 introduces, itemise: £120 million for the AI Security Institute, £70 million for the NCSC centre, £24 million for public sector AI defence, £60 million for Cyber Shield and £40 million for the National Agentic Security Unit. That totals £314 million. Row 31 of the Treasury's own Table 6.1 policy-decisions scorecard shows an "Artificial Intelligence" line costing £490 million over six years, but its own title defines that money as covering three things: AI security, "accelerate adoption across the economy", and "support strategically important UK AI capabilities" [3]. The Budget's own paragraph 4.134 introduces the adoption and capabilities spending separately, after 4.129's security section, under the heading "Unlocking the opportunities of AI" [2] | The "around £500 million" security figure is arithmetically real only if it includes roughly £176 million that the same document, a few pages later, says is for a different purpose |
A unit with no defined target
Paragraph 4.131 sits inside a section headed "To deliver a step change in UK cyber defence, the government will", and it reads as a list of concrete commitments [2]. Three of its four bullets are genuinely specific: the National Exposure and Real-Time Vulnerability Centre sits at the National Cyber Security Centre with a stated job, understanding and reducing national cyber exposure; the public sector AI defence programme hardens a named estate, local authorities, hospitals and schools; and the Cyber Shield funds a first phase of agentic cyber defence capability in partnership with industry [2]. The fourth bullet, the National Agentic Security Unit, is the only one of the four without a host organisation named and without a description of what it is actually looking for beyond "rogue agentic systems" [2]. Paragraph 6.106, in the Budget's later policy-decisions chapter, restates the unit in almost identical but not quite matching terms: "world-first capabilities to detect and respond to misaligned agentic activity" [2]. "Rogue agentic incidents", "rogue agentic systems" and "misaligned agentic activity" are three different phrases for what is presumably one thing, and none of them is defined anywhere in the 166-page document, its policy costings annex, or its Table 6.1 and 6.2 policy decisions spreadsheet, searched directly for each term. Nor does the Budget say whether the unit has any legal power to compel information from a company whose AI agent it wants to investigate, who it answers to, or when it will exist. The Cyber Security and Resilience Bill, reaching Third Reading in the House of Lords this week, gives GCHQ's own confirmations a defined statutory role and a six-month deadline, imperfect as that deadline is [4]. The National Agentic Security Unit gets none of that: it is a name, a figure and a one-sentence job description.
The £500 million that is also someone else's £176 million
Paragraph 4.129's "around £500 million" is the sentence that frames the whole section: it is the number a reader carries away as the cost of making AI secure. The two paragraphs it directly introduces, 4.130 and 4.131, name five costed items totalling £314 million, plus two bullets with no stated price, locating AI security expertise outside London and a future consultation on liability for insecure technology [2]. That leaves £186 million unaccounted for in the prose itself. The Treasury's Table 6.1 scorecard, the document that actually scores each Budget measure's cost to the Exchequer year by year, does contain a line that gets close: row 31, "Artificial Intelligence", costing £35 million, £115 million, £185 million, £145 million, £5 million and £5 million across the six years from 2026-27 to 2031-32, a total of £490 million [3]. But that line's own title is explicit that it funds three things, security, "accelerate adoption across the economy", and "support strategically important UK AI capabilities", not security alone [3]. The Budget's own structure treats adoption and capability-building as a later, separate stage: paragraph 4.134 opens "With these strong foundations in place, the UK can capture the benefits of AI", introducing a new heading, "Unlocking the opportunities of AI", and a fresh set of commitments, £30 million for local AI partnerships, a £60 million AI Pathfinder Fund, an additional £23 million for the AI Economics Institute [2]. The document presents foundations and opportunities as sequential and distinct. The only Exchequer accounting line that produces a figure close to "around £500 million" does so by treating them as one.
Two documented mistakes, four stars. (1) The £40 million National Agentic Security Unit is given an investigative and response remit over "rogue agentic incidents", "rogue agentic systems" and "misaligned agentic activity", three phrases used interchangeably across the document, none of them defined anywhere in the 166-page Budget, its policy costings annex or its policy decisions spreadsheet, and the unit is given no stated power, no named accountable body, and no date by which it must exist. (2) The Budget's claim that "around £500 million" will "build the secure foundations for an AI-enabled economy" is £176 million more than the £314 million its own itemised breakdown in the same two paragraphs actually costs, and is only matched by the Treasury's separate scorecard, whose own title for the matching line states it also funds AI adoption and capability-building, spending the Budget's own later pages treat as a distinct, subsequent stage.
Credit where due
The section is not a black box. Four of the five costed items in paragraphs 4.130 and 4.131, the AI Security Institute top-up, the NCSC exposure centre, the public sector AI defence programme and the Cyber Shield, are each given a specific figure and a specific host, more granular disclosure than many Budget technology announcements receive. The fourth bullet's commitment to "consult next year on how to tackle insecure technology at source and consider where liability for cyber breaches should sit" is honestly flagged as unresolved future work, not dressed up as a decision already taken [2]. And the Treasury keeps at least one adjacent AI funding stream, the Sovereign AI Fund, on its own separate scorecard line rather than folding everything AI-related into a single number, which is why the £490 million row 31 discrepancy is visible at all rather than buried further [3].
Sources
- Budget 2026 (HTML), HM Treasury, 28 October 2026.
- Budget 2026 (PDF, HC 726), HM Treasury, 28 October 2026.
- Table 6.1 and 6.2 Budget 2026 Policy Decisions (spreadsheet), HM Treasury, 28 October 2026.
- HL Bill 62 (as amended on Report), Cyber Security and Resilience (Network and Information Systems) Bill, UK Parliament Bills API, 26 October 2026.