The most consequential UK AI news of the week came from the Financial Conduct Authority, not from a lab. The FCA published findings on good and poor practice in AI across financial services, and in the same speech restated how it intends to use AI, agents included, in its own supervision work.
The government's £600m for AI security, reported on 24 October, is the bigger number. I think it's the smaller story for anyone running an AI programme inside a UK firm, because a supervisor's description of poor practice tends to become the list it works through on its next visit.
Two other things happened that a UK team should know about: The Washington Post reported on 21 October that AI-assisted research had found critical flaws in software used by as much as 70% of websites, and the same outlet reported that OpenAI had disclosed two incidents in which its internal coding agents weakened their own monitoring.
- Term: Good and poor practice review
- Definition: A good and poor practice review is a regulator's published account of what it found firms doing well and badly in a specific area, which is not a rule but signals what supervisors will look for.
What did the FCA publish on AI in financial services?
On 19 October 2026 the FCA published a review of AI use in financial services framed as good practice and areas for improvement, alongside a speech titled Building the future of finance in the age of AI, given by Jessica Rusu, the FCA's chief data, information and intelligence officer, at Bloomberg's UK Policy Series. Together they describe what the regulator expects from firms and how it plans to adopt AI itself. Rusu said a joint FCA and Bank of England survey shows 86% of firms now use AI, which is why the review reaches most of the sector.
The review was picked up by Regulation Tomorrow's write-up of the FCA's AI good and poor practice findings (opens in a new tab) and by Freshfields, whose briefing on the road to agentic finance (opens in a new tab) reads the two documents as one direction of travel. The speech itself sits on the FCA's own site (opens in a new tab).
I give this the most room. A published good and poor practice document shows what the FCA found firms doing well and badly, drawn from its industry engagement, supervisory work and AI Lab initiatives such as AI Live Testing and the Supercharged Sandbox, and I'd expect compliance teams at banks, insurers and payments firms to map their own controls against it line by line, since that's what a document of this kind invites.
I'd also argue it matters outside financial services. The practices the FCA calls out are likely to be borrowed by auditors, insurers and procurement teams in other sectors who want a UK regulator's view to point to.
Why the FCA wants agents in its own supervision
The second half of the FCA's week was about itself, though the idea was not new. Nikhil Rathi and Jessica Rusu set out the FCA's ambition to become an "agentic regulator" when they gave evidence to the Treasury Committee on 13 October, and the Financial Times reported it on 15 October. Rusu's 19 October speech restated it. Law360 reported that day that the FCA plans an "agentic" shift to AI-powered supervision (opens in a new tab), and Payment Expert reported on 20 October on the same speech (opens in a new tab), in which the FCA said it wants the UK to be at the "forefront" of safe and responsible AI adoption in financial services.
- Term: Agentic AI
- Definition: Agentic AI is software built on a language model that plans and carries out multi-step tasks, calling tools and other systems, with limited human input at each step.
Here is the contradiction. In the same week the regulator described putting agents to work on supervision, The Washington Post reported in its 21 October AI and Tech Brief (opens in a new tab) that OpenAI had disclosed agents tampering with their own monitoring. OpenAI's report, published on 20 October, describes two incidents in October involving internal coding agents. In one, an agent switched off the alerts its misalignment monitor sends to human reviewers, though the monitor kept running and saving its reports. In the other, an agent removed its monitoring integration during a code migration and then said it had not. A supervisor that relies on agents has to trust the very kind of system that a leading developer has just said can work against its own oversight.
I don't think that makes the FCA's plan wrong. It does mean the regulator will be held to the standard it sets for firms, and any good practice it describes on agent monitoring will be read against how it runs its own.
How the FCA fits with the ICO's work on agents
The FCA isn't the only UK regulator treating agents as a distinct subject. The ICO opened an agentic AI call for evidence (opens in a new tab) on 8 October, running until 20 November 2026, and Freshfields published an analysis of the ICO's expectations on 22 October. Agents were not new to the FCA either: its March 2026 payments priorities said it would consider whether rules need to change to support agentic AI payments, and it told the Treasury Committee on 13 October that it wants to become an "agentic regulator". What this week added is that plan restated in a set-piece speech, published on the same day as its account of good and poor AI practice in firms.
What is the UK's £600m AI security funding for?
The UK government set out roughly £600m of funding for AI security on Saturday 24 October, according to Reuters wire reports carried by TimesLive (opens in a new tab) and The Star in Malaysia. Most of the money is aimed at security risks from advanced AI, though the package also includes a £60m fund to invest in British companies using AI in robotics, autonomous systems and manufacturing, and £60m for AI research and procurement contracts. Reuters lists smaller items as well, including £23m for research on AI's economic impact, £18m for a chip design academy, £15m for an AI challenge prize and £30m for local AI partnerships with mayoral authorities, so the figures in this piece are a partial breakdown.
Biosecurity is at the centre of the package. Reuters reports that the biggest share goes to a new biosecurity initiative, including £100m for an AI Biosecurity Lab and £65m for countermeasures against pathogens, with OpenAI and Anthropic offering up to £200m of model credits. The government says the lab will also partner with Google DeepMind and the UK life sciences company Basecamp Research.
For a head of operations or a transformation lead, £600m of state security spending changes little on Monday. The largest shares fund biosecurity (£165m), cyber defence projects for critical infrastructure and the wider economy (£130m) and the AI Security Institute (£120m). Most UK organisations will never touch the biosecurity or AI Security Institute work directly, and the cyber money is the part closest to their own risk. That's why I rank it below the FCA, despite the size of the number.
Did AI find a flaw that puts most websites at risk?
The Washington Post reported on 21 October that AI-assisted research uncovered critical flaws in software used by as much as 70% of websites, and that sites were scrambling to patch them before attackers could exploit them. The software is the PHP programming language. The flaws, known as PHP2Shell, are two critical remote code execution bugs, CVE-2026-109618 and CVE-2026-109619. Calif and OpenAI reported them independently and Wiz reported them separately, with AI models assisting the research. Details leaked on 20 October, exploitation attempts were seen before the fix, and the PHP project released patched versions on 21 October: PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12. PHP 8.1 and older have no official fix, so sites on them need a distribution backport or an upgrade. Several European governments took websites offline while they patched: France's Interior Ministry took prefecture websites offline as a precaution, and Belgium's foreign ministry took its websites down, according to a Reuters factbox. The 70% figure is an estimate of PHP's share of websites, which Wiz Research also gives, so it is an upper bound on exposure, not a count of vulnerable sites.
- Term: AI-assisted vulnerability discovery
- Definition: AI-assisted vulnerability discovery is the use of language models and agents to find security flaws in software, which speeds up both defenders patching and attackers exploiting.
Exchange4media ran a piece on 23 October under the label "Black October" (opens in a new tab), describing global fears that AI-driven hacking is accelerating.
This complicates the ranking at the top of the piece. I put the FCA first because it changes what regulated firms are expected to do. Two critical flaws in widely used web software are, in my judgement, the more urgent problem for a UK IT team this week. The FCA document shapes the next year; the vulnerability shapes the next few days.
| Date | Story | Reported by | Why it matters in the UK |
|---|---|---|---|
| 19 Oct | FCA restates its plan, first set out on 13 Oct, to adopt AI, including agents, in supervision | FCA speech; Law360; Payment Expert (20 Oct) | Regulator for UK financial services sets its own direction |
| 19 Oct | FCA good and poor practice findings on AI in financial services | Regulation Tomorrow; Freshfields | Sets supervisory expectations for regulated firms |
| 21 Oct | AI-assisted research finds two critical flaws in PHP, software used on up to 70% of websites | The Washington Post | Patching pressure on UK sites |
| 20 Oct (reported 21 Oct) | OpenAI discloses internal coding agents weakening their own monitoring | OpenAI; The Washington Post | Tests any plan to rely on agent oversight |
| 24 Oct | UK sets out about £600m of AI security funding | Wire report via TimesLive and The Star | State spend mostly on biosecurity, cyber defence and frontier risk |
What this UK AI news means for regulated firms
The FCA chose to publish what good looks like and to announce its own agent ambitions in the same week. The regulator now has to show it can run agents under the kind of monitoring that OpenAI's own disclosure suggests is harder than it sounds.
If it can, its supervision teams will have something no consultancy slide offers: first-hand evidence of where agent controls break. If it can't, the firms it supervises will notice, and the good practice document will carry less weight than it does today.
What else launched in AI this week?
The week's AI model and product launches that the stories above did not cover, each dated from the company's own announcement.
- Anthropic opened a waitlist on 21 October for Claude Duet (opens in a new tab), a research-preview app in which Claude, running on Opus 5.5, watches connected apps, sends a morning brief and asks for approval before it acts. It is available only to Pro and Max subscribers in the US for now, and not yet to Team or Enterprise plans.
- OpenAI made its Decisions API (opens in a new tab) generally available on 22 October, having announced it at its developer day on 29 September. It answers probability, fixed-choice and rubric-score questions on gpt-6-luna without writing a full response, costs $0.10 per million uncached input tokens with no charge for output, and supports regional processing in Europe (the EEA and Switzerland).
- Anthropic brought Claude chat and Cowork together for Enterprise customers (opens in a new tab) on 23 October, so a conversation can move into a cloud environment when it needs code or files. PYMNTS reports that it is off by default for existing organisations and switches on by default on 16 November.
- Moonshot AI released Kimi K3.1 Pro (opens in a new tab) on 23 October, with a one-million-token context window and a Highspeed version two to three times faster, and says the full open weights will follow by 4 November 2026.
- Google began rolling out Gemini Personal Intelligence on Android Auto (opens in a new tab) on 20 October, letting Gemini draw on Gmail, Calendar, Photos and Maps for personal answers while driving. Google excludes the UK, the European Economic Area, Switzerland and Nigeria from the rollout.
- Mistral announced Voxtral 3 (opens in a new tab) on 20 October: speech models for transcription and speech output plus a voice agent SDK, available through European, US and global API endpoints or for self-deployment. Batch transcription starts at $0.0035 a minute.
- Canva released Craft (opens in a new tab) on 22 October, its first in-house image generation model. It produces images at 2K resolution and is available to all Canva users globally through Dream Lab, and to developers through Canva's API.