Reading time ~15 mins

The Dutch Cabinet's AI vision leaves finance off its priority list and warns of synchronised markets, and its promised cyber programme names neither banks nor the Dutch central bank. An ECB board member sets banks' talk of AI savings against rising IT and cyber incidents. Revolut swaps Claude for self-hosted Chinese models, Deutsche Bank reports AI projects reaching production in as little as three months, and OpenAI limits a cyber model to vetted defenders. Google describes criminals reselling hijacked AI access, and NIST records how incident plans fall short for AI.

Regulatory

An ECB board member sets banks' investor talk of AI savings against a near doubling of IT and cyber incidents. Authority

Philip Lane, a member of the European Central Bank's Executive Board, gave the keynote AI adoption in Europe at the Federal Reserve Bank of San Francisco on 30 October. Most of it is macroeconomics, but one section reads bank data. The ECB ran 761 earnings-call transcripts from 51 euro area banks through an AI model. Executives increasingly raise their AI plans with investors, and almost half of those references concern cost optimisation and automation. Mentions of cyber operational risk "remain more limited", Lane says, in contrast with the evidence of rising threats. In a sample of large-loss events from ECB supervisory reporting, IT, cyber and related incidents were 10-12% of operational-risk events in 2025, nearly double 2018. That count also takes in system failures and customer fraud, and Lane says the share due to AI alone is hard to isolate. Operational-risk losses stayed contained at 0.9% of aggregate capital, though concentrated in tail events. Loans to euro area AI companies are growing faster than other lending, at just under €200 billion. Lane speaks as an economist and announces no supervisory step. The speech came a day before the 31 October deadline for banks' action plans on AI-enabled cyber threats, which this brief covered on 2 November. It puts the ECB's own data beside what bank leaders tell investors, so any analyst can ask a euro area bank the same question at full-year results.

European Central Bank: AI adoption in Europe, keynote by Philip R. Lane

The OCC framework that ran here on 5 November counts a running but wrong agent as a resilience problem. Authority

On 5 November this brief led with the examiner framework of the Office of the Comptroller of the Currency (OCC), the US national bank regulator. Dated 4 November, it covers what the OCC calls Super Intelligence and this brief calls AI. Nothing has changed since. It still sets no expectations for banks, so any use is a bank's own choice. The full text does go past the kill switches readers had from the Financial Stability Board on 16 October. Section 2.10 counts incorrect operation as a resilience problem alongside loss of service, because an agent "could remain available while taking inappropriate or harmful actions". When a misfire affects customers, a bank may set up a separate workstream to identify the complete affected population, assess individual harm, correct it and independently verify completion. Section 2.15 says audit may assess whether a backstop, such as a control agent, a circuit breaker or a human process, is independently governed, tested and effective. Section 3.1 puts a fully autonomous, customer-facing payment agent in the top tier regardless of how well it performs in testing. A footnote says the framework was itself prepared with the technology it covers and approved by a human, with no detail on either. An availability report would not show such an agent. The harder test the text implies is whether a bank could name every customer it touched.

OCC: Supervisory Framework on Bank Use of Super Intelligence

Perspectives

Jakub Szarmach praises a control matrix for naming who owns each control, but the supplier writes the customer's part. Corporate

Jakub Szarmach, who runs a library of AI governance resources, wrote on LinkedIn on 4 November about the Cloud Security Alliance's AI Controls Matrix, a voluntary set of 247 controls with a 320-question supplier questionnaire. Jakub Szarmach writes: "AI controls are only useful if you know who actually owns them." The matrix splits each control across five roles, from model provider to AI customer. His post leaves out who writes that split. The supplier fills in the questionnaire, the customer column is optional, and the Alliance offers customers no template of their own. Its registry showed 102 organisations on 6 November at Level 1, which has no outside audit, and six at Level 2, and the Alliance runs that programme. UK supervisors found on 27 October that most third-party AI is taken on the vendor's word. A questionnaire gives a bank better questions and the same self-declared answers, until the customer's part is settled in the contract.

Cloud Security Alliance: AI Controls Matrix v1.1 (shared by Jakub Szarmach)

Tony Moroney argues that AI customer service can cut a firm's work by moving it onto the customer. Advisory

Tony Moroney, who writes on digital change in financial services, wrote on LinkedIn on 5 November that a self-service session can end well on the dashboard. The customer then contacts another channel and starts again. Tony Moroney writes: "The individual systems may have performed exactly as designed, but the customer has become the integration layer between them." A lower cost per interaction, he says, may reflect work moved and not removed, and some of it may never appear in service data. He gives no figures. Cost per interaction stops at the end of the session, so the repeat contact and complaints it misses can land in another team's numbers.

Tony Moroney via LinkedIn (LinkedIn; original source not verified)

Netherlands & Sovereignty

The Dutch Cabinet's AI vision leaves finance off its priority list and warns of synchronised financial markets. Authority

On 6 November the Cabinet sent parliament its AI vision, Kiezen voor AI voor iedereen, drawn up under State Secretary Willemijn Aerdts. It chooses open strategic autonomy in a strong EU, and calls full independence neither realistic nor desirable for the Netherlands. Not every sector can be a national priority, it says. It names health and care, energy, food production, mobility and logistics, security and defence, and the technology industry. Financial services is not among them. Finance appears instead as a sector where the AI Act should give confidence and reliability is a competitive advantage. The annex does count knowledge-intensive business services among Dutch strengths. The Cabinet makes €120 million available for Dutch participation in a joint European AI project, and leaves the financing of AI infrastructure to the market. Advice on a possible Dutch AI safety institute is expected in December, and letters on AI infrastructure and data centres follow in the first quarter of 2027. The vision creates no new legal duty for a bank.

The annex holds the passage closest to a bank. It says the limited supply of AI and cloud providers raises the risk of more synchronised behaviour in financial markets, through similar models and data sets. Similar orders executed at the same time could raise the chance of global market instability. This is one illustration in a section on national security, and it names no supervisor, measure or date. Set beside the priority list, it puts finance among the named risks and outside the named opportunities.

The same day brought the cyber action programme this brief said on 28 October was due in November. The Minister of Justice and Security sent it with the Cybersecurity Assessment Netherlands 2026, which shows for the first time that attackers demonstrably use AI in cyberattacks, mostly in a supporting role. The programme says the question is no longer whether but when the Netherlands faces a society-disrupting cyber incident. It runs to the end of 2027 within existing budgets and includes work on faster and automated patching. Neither the programme nor the Minister's letter mentions banks or De Nederlandsche Bank (DNB), the Dutch central bank. On 6 October the Tweede Kamer asked the government to have DNB test how prepared the financial system is for AI attacks on transfers, payments and markets. The letter closes a list of motions, and that one is not on it. Finance has its own regime under DORA, the EU's digital resilience law, and a separate answer may still come.

The Cabinet also replied to the Social and Economic Council on AI and work. It is considering updating the Works Councils Act to cover strategic developments in the enterprise, with a Council proposal to come. It also notes signals that employers name AI as a reason for layoff rounds. Nothing is decided. If it goes ahead, a bank's AI plans would reach its works council earlier than they do now.

Rijksoverheid: Kabinet kiest voor AI die Nederland vooruithelpt | Tweede Kamer: Kabinetsvisie op Artificial Intelligence | Tweede Kamer: Actieprogramma Toekomstbestendige Digitale Weerbaarheid | Tweede Kamer: Cybersecuritybeeld Nederland 2026 | Tweede Kamer: letter of the Minister of Justice and Security | Tweede Kamer: Kabinetsreactie op het SER-advies over werk en AI

Industry & competition

Revolut says self-hosted Chinese models cut AI coding costs by up to 90%, and Australia lists what moves in-house. Media

TechRadar reported on 4 November that Revolut, the London-based digital bank, has moved coding workloads from Anthropic's Claude to self-hosted open-weight models from China's Z.ai. Open-weight means the trained model can be downloaded and run on hardware the buyer chooses. Inference costs, the cost of running a model, reportedly fell by as much as 90% in less than a year. "The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," infrastructure lead David Abitbol said. He also claimed 70% better lead-time productivity and 60% better coding-time productivity, with no baseline or method. The 90% covers inference only. Revolut reportedly runs 56 Nvidia B200 chips, and no total is given for hardware, power, upkeep or staff. This is a trade outlet relaying conference remarks first reported by The Register.

The Australian Signals Directorate, that country's cyber agency, published a guide to operating open-weight models on 5 November that lists the other side of the ledger. Safeguard upkeep, abuse detection, security updates, monitoring and incident response "do not exist in a self-hosted deployment unless the organisation builds and maintains them", it says. It asks for a register of approved models, each with an accountable owner and a next review date. Country of origin is one element of the assessment, it says, and not by itself evidence that a model is trustworthy or untrustworthy. Where provenance cannot be established to a level suited to the use, the model should not be adopted. The guide binds no one in the EU. On 22 October bunq described moving its support assistant to open-weight models hosted in the EU, after testing a Chinese-made one at about 60% of the cost. Two digital banks active in the Dutch market have now put a price on this trade in public. Neither has published what the agency's list would cost.

TechRadar Pro: Revolut reveals it cut AI coding costs by cutting out Claude entirely | Australian Signals Directorate: Selecting, deploying and operating open-weight AI models securely

Deutsche Bank says moving AI controls to the start cut idea-to-production time from 12 months to as little as three. Corporate

Divya Veeramahalingam, Deutsche Bank's AI platform product lead, wrote on the bank's site on 6 November that projects on a standard route went from around 12 months to as little as three. Teams get staged workflows, approved development environments, approved data routes, reusable connectors and technical blueprints. An AI Access Layer gives applications and agents common access to bank-approved models, with access controls and oversight of usage and cost built in. "As little as" marks a best case. The post gives no project count and no spread, and it cannot separate the route from the choice of projects. A FinTech Futures piece the same day by practitioner Sandi Samantaray argues the same point. A reusable approval route, it says, can turn bespoke review from months into weeks, with each model change tested against a fixed pack of past cases. Both are self-reported. A large EU peer now states that moving controls forward bought the speed, which gives a board a number to ask for, idea-to-production time per project.

Deutsche Bank: Trust by design, how responsible AI speeds up innovation | FinTech Futures: What it takes to ship AI inside a regulated bank

Capgemini's survey finds over half of banks and insurers would let the big cloud providers govern their AI agents. Advisory

The Capgemini Research Institute released its World Cloud Report for Financial Services 2027 on 5 November, from a poll of 1,100 financial-services leaders in 14 markets in June and July. In its press release, 52% of banks and insurers name hyperscalers, the largest cloud providers, as their preferred partner to govern multi-agent deployments, in which several AI agents work together on one process. It says 59% are piloting or deploying such systems, which lumps pilots with production, and 10% have enterprise-wide orchestration. Roughly half cannot consistently show a return on agentic AI. Capgemini sells cloud work, the question wording is not given, and the survey measures stated preference, not contracts. Readers have had the adoption gap from several surveys. The preference is the new number. Governance placed with the firm that hosts the agents is harder to move than the workloads themselves, which sits awkwardly beside the Cabinet's wish, published the next day, to reduce risky dependence.

Capgemini: Financial institutions accelerate transition to multi-agent AI in drive for bottom-line value

Innovation

OpenAI limits GPT-6 Cyber to vetted defenders and says it found 81 unknown high-severity Firefox flaws. Vendor

OpenAI released GPT-6 Cyber on 5 November to a select group of approved customers of Daybreak Red, its restricted cyber tier, with wider access planned. It reports 48.5% on its ExploitGym test against 32.1% for GPT-5.6 Cyber. It says its researchers used the model to develop 948 Linux exploits and identify 81 previously unknown high-severity flaws in the Firefox web browser. These are OpenAI's own figures, with no independent check of the Firefox claim. Its system card rates the model Critical in cybersecurity, the highest rating in its framework and the one this brief reported for GPT-6.1 Terra on 30 October. Access requires identity checks, multi-factor login and an approved list of network addresses. One term reaches a bank's data protection team. Customers on Zero Data Retention, a contract under which the vendor keeps no prompts, generally must accept Private Safety Processing. That is automated review of related interactions, run on encrypted records with no staff access to the text, and the customer holds the records for 30 days. So no retention here does not mean no review. On 2 November this brief carried the ECB's Pedro Machado calling patching banks' biggest challenge, and the Dutch programme published on 6 November has an action on shrinking patch time. A model that finds flaws at this rate, if the claim holds, shortens the time a bank has to fix them.

OpenAI: Introducing GPT-6 Cyber | OpenAI Deployment Safety Hub: GPT-6 Cyber system card

Mistral adds managed agents to a platform BNP Paribas says it runs on its own infrastructure. Vendor

Mistral announced managed agents and enterprise apps for its Studio platform on 5 November, with added workflow, monitoring and evaluation features. Studio can run in Mistral's cloud, a customer's private cloud or on premises, and teams can limit tools and data and insert human approvals. Mistral names ABANCA and BNP Paribas among production users. Su Yang, BNP Paribas's head of AI and IT innovation, says in the post: "we run Mistral Studio on our own infrastructure and keep those systems under our control". Neither bank says it uses the new features, and the post gives no outcomes. Crédit Agricole's group platform with Mistral ran here on 23 October. With BNP Paribas's statement, a self-hosted European route for agents now has a named large-bank user, which turns the Cabinet's question about acceptable dependence into a live sourcing option.

Mistral AI: Build Enterprise Agents, Workflows and Apps in Mistral Studio

Security

Google describes criminals hijacking firms' AI access across providers, with charges reaching $1 million at some organisations. Vendor

Google's threat intelligence group and Mandiant, its incident response arm, published a study on 3 November of what they call LLMJacking, in which criminals take over a firm's paid access to AI models. It draws on multiple 2026 incidents across several cloud and AI providers, in the Americas and Asia Pacific, in sectors including financial services. Charges exceeded $100,000 in a single day in individual incidents and reached $1 million at some organisations before access was revoked. The usual way in was exposed or stolen cloud login tokens, an ordinary identity failure and no flaw in an AI model. One group tested stolen keys with minimal requests that stayed below volume alerts. Access is then resold through what Google calls LLM Mixers, illicit gateways that pool and meter stolen model access, and one service advertised an unlimited plan at about $60 a month. The same intrusions included credential theft and copying of code repositories, so the bill may be the visible part of a wider break-in. Google sells security services and names no victim. Its advice includes spending caps, quotas per model and phishing-resistant login. On 26 October Nate B. Jones argued that a stolen AI key should be treated like stolen cash, and these are the first-hand figures. The loss arrives as a cloud invoice, so it may reach the owner of the cloud budget before any fraud or security team.

Google Cloud Blog: The Rise of LLMJacking

NIST's workshop report records that a secure AI system can still cause an incident, and evidence must be prepared beforehand. Authority

The US National Institute of Standards and Technology (NIST) published IR 8621 on 2 November, a report on its 14 May workshop on incident response for AI. NIST writes super intelligence, a label directed by Executive Order 14434 with no technical definition, so this brief writes AI. The report records speakers' views and sets no requirement, private firms may use it voluntarily, and the word bank does not appear. It separates attacks on AI systems from harm the systems cause through misuse or malfunction. A recurring point was that a technically secure system can still cause an incident through wrong outputs or unacceptable decisions. Robert Saul of Amazon Web Services said containment can itself do damage. Killing an agent can corrupt downstream data, and rolling back a model can break dependent services, so he recommended decision trees naming who may approve each step. Another speaker proposed fixing a record of model versions, prompts, tools and retrieved context before any incident, because a model that changes and answers differently cannot be replayed later. A CrowdStrike speaker suggested 30 days as a baseline for keeping prompts, while the report notes such logs may hold personal data. Several speakers sell these capabilities, and no bank or financial supervisor spoke. NIST's next steps are headed possible and carry no dates. They include updating its incident-response guide, with a possible part on misuse and malfunction. The OCC text above reaches the same point from the supervisor's side. A bank's incident process starts from a failed system, so a wrong answer from a healthy one may reach complaints or model monitoring first, and the report does not say who owns it.

NIST: IR 8621, NIST Workshop on SI Incident Management

On the radar

• India's national cyber agency CERT-In warned on 4 November that a campaign it calls CRAFTY FALCON is using AI-enabled tooling against non-bank lenders, fintechs and payment processors, with activity returning after incomplete clean-up. CERT-In