Reading time ~12 mins

The Dutch cabinet says organisations' cyber defences are not keeping pace with AI and promises an action programme in November. Its answer does not take up parliament's call for a DNB test of AI attacks on the financial system. A survey finds about one Dutch board in five discusses AI's opportunities or risks, and government analysts set out when reliance on one tech provider stops being acceptable. OP Pohjola and BNP Paribas report assistant results that do not compare, and Itaú buys the startup that built AI models on its own data. OWASP finds no AI attack agent finished an exploit without a person, and a study of one firm shows a change of reviewer hiding an agent's real improvement.

Regulatory

The Dutch cabinet says organisations' cyber defences are not keeping pace with AI, and promises an action programme in November. Authority

On 26 October State Secretary Aerdts of Economic Affairs and Climate, also answering for the Minister of Justice and Security, replied to written questions from MP Stoffer. The questions concerned warnings from researchers at leading AI companies. The cabinet says it shares the researchers' concern that the most capable AI models are developing faster than the safety and control measures around them. It points to the July case in which OpenAI's test agents gained internet access and hacked the code platform Hugging Face. AI, it says, lowers the threshold for cyberattacks and raises their speed, scale and efficiency. The threat landscape is moving fast, and the digital resilience of Dutch organisations and businesses is not growing at the same pace. The cabinet is therefore preparing an action programme that adds to the Dutch Cybersecurity Strategy 2022-2028, aimed at cyber threats including those from advanced AI. The Minister of Justice and Security will present it to parliament in November 2026. It arrives together with the strategy's progress report and the Cybersecurity Assessment Netherlands 2026, the government's annual threat picture. The cabinet keeps the EU AI Act as the legal framework, while noting that the EU AI Office can act only once a model is placed on the EU market. It does not support a unilateral pause in AI development or a single all-embracing global AI treaty. It does leave room for further measures, far-reaching restrictions among them, as evidence develops. It also warns that computing power and capital concentrated in a few commercial firms can, in extreme cases, create risks that disrupt society, such as large-scale cyberattacks.

The answer does not mention banks, DNB or the financial sector. On 6 October the Tweede Kamer, the Dutch lower house, adopted a motion by MPs Dassen, Stultiens and Van Brenk. It asks the government to have DNB test how prepared the Dutch financial system is for AI attacks on bank transfers, payment systems and markets. The cabinet's answer does not take that request up. The November programme is the next dated government document in which it could. It comes about a month after 31 October, the date by which the ECB asked every bank it supervises for an action plan on AI-enabled cyber threats. A motion places no legal duty on a bank, and the programme's content is not yet known.

Tweede Kamer: Antwoord op vragen over risico's van geavanceerde AI

Netherlands & Sovereignty

A survey of 255 Dutch directors and supervisors finds about one board in five discusses AI's opportunities or its risks. Corporate

The Nationaal Register, which trains and recruits board members, and VNO-NCW, the Dutch employers' federation, published the third edition of their board survey on 21 October. They ran it with NLdigital, the Dutch digital industry association, and CIO Platform Nederland. It asked 255 executive and supervisory board members across Dutch organisations. Three quarters say digitalisation is strategically important to their organisation. More than 70% say their organisation is considering or already building new business models because of digital change. AI trails well behind. It is a strategic agenda item in 46% of boardrooms, and 4% of organisations say they have integrated it into strategy. 19% say their board discusses the opportunities of AI, generative AI included, and 22% say it discusses the risks. More than three quarters see AI mainly as a way to make existing processes more efficient. The two organisations call on directors and supervisors to build their own understanding of AI and treat it as a matter for the whole board, not one for the CTO or CIO alone. The survey gives no figures by sector, so banks, which face supervisory expectations on AI oversight, may sit above this average. It is still the first Dutch cross-sector baseline for a count any board office can make from its own minutes. That count is how often AI's opportunities and its risks reach the full board as items in their own right.

VNO-NCW: Aandacht voor digitalisering groeit, strategische blik op AI blijft achter

The government analysts behind the national risk analysis set out when reliance on one tech provider stops being acceptable. Authority

On 27 October this brief carried the 2026 national security risk analysis by the Analistennetwerk Nationale Veiligheid, the government's network of risk analysts. One of its scenarios has public bodies drifting into dependence on a single large technology provider. A 32-page companion report on the dark sides of Big Tech, published by RIVM, gives the reasoning behind that scenario. The analysts call the scenario unlikely today and use it to make long-term weak points concrete. Outsourcing to private providers, the report says, need not create an unwanted dependency as long as meaningful control over the process remains. It names two ways that control is lost. One is services built on algorithms whose choices are no longer transparent. The other is lock-in, when one provider's services fit well together but work poorly with anyone else's. Among warnings from several sectors, it cites the 2025 DNB and AFM report on the financial sector's digital dependence. It also gives a case from The Hague. After US sanctions on the International Criminal Court, US companies placed restrictions on several court staff in 2025 and 2026, including access to bank accounts and to Amazon, Google and Microsoft services. Its recommendations move the focus from business continuity to the continuity of society, and accept long-term choices that may look less economical in the short term. Among them are European cloud, data and AI capacity and open standards that prevent lock-in. The report is written for government and places no duty on a bank. Its test asks whether control is kept, not whether a service is outsourced. That is close to the logic of DORA, the EU's digital resilience law, which asks banks to keep control of critical providers and a way out. The court case shows a provider's home state reaching an organisation in the Netherlands through its suppliers, bank accounts included.

Analistennetwerk Nationale Veiligheid: Themarapport De duistere kanten van Big Tech (publication date unverified)

Industry & competition

OP Pohjola and BNP Paribas publish results for their customer assistants, measured in ways that do not compare. Corporate

OP Pohjola, the Finnish cooperative banking and insurance group, published its interim report for January to September on 27 October. In his review, chief executive Timo Ritakallio reports that OP Aina, its AI assistant for banking and insurance, had almost 8 million customer encounters in the nine months, up 36% on a year earlier. It handled 3.9 million of all customer contacts and answered 22% of the queries customers made across service channels. A record 86% of feedback on it was positive. OP says it also uses AI in insurance claims, work against financial crime and support for staff, and it published an AI statement in September. The same day BNP Paribas described EVA, the assistant of its consumer-credit subsidiary FLOA, which has run for two years on the website, by phone and in the app. BNP Paribas says EVA handles 78% of the contacts it receives without an adviser, with a Net Promoter Score of +44 from nearly 15,000 responses. Customers can ask for an adviser at any time, and a person always handles complex transactions. Both sets of figures are self-reported, and neither firm says how it counts a query as answered. The two rates measure different things. OP's 22% is a share of all customer queries across its channels. FLOA's 78% is a share of the contacts that reach EVA, in a business built on instalment and split-payment credit. A board that sets its own assistant against either figure would be comparing a reach measure across the whole bank with an automation rate inside one channel. OP now reports the number in its quarterly results, next to its claim that AI frees staff for the advice customers value most.

OP Pohjola: Interim Report 1 January to 30 September 2026 | BNP Paribas: AI at BNP Paribas #5, EVA

Itaú Unibanco agrees to buy NeoSpace, the startup that built AI models on the bank's own data, for about $589 million. Media

Itaú Unibanco, Latin America's largest bank, has agreed to buy all shares in NeoSpace A.I. Technologies for 2.9 billion reais, about $589 million. Retail Banker International reported the deal in an article Yahoo Finance carried on 27 October. NeoSpace builds foundation models, large AI models trained on broad data that other applications are built on. It has already built models on Itaú's proprietary data, which the bank uses in financial management, credit, customer relationship management and fraud prevention. Itaú plans to apply the startup's capabilities across its products and services, including its generative AI assistant, ia.i. NeoSpace will keep operating independently with its current staff. Itaú pays part of the price on completion and the rest in instalments over five years, subject to agreed conditions. The deal needs approval from CADE, Brazil's competition authority, and from the Central Bank of Brazil. This month Crédit Agricole's new group AI company signed a partnership with Mistral, and Santander signed a three-year framework with NVIDIA. Itaú goes a step further and buys the model maker whose work already runs in its credit and fraud decisions. Paying over five years on conditions, with the team kept in place, suggests it is buying the people who build the models as much as the models themselves.

Yahoo Finance: Itaú Unibanco gains full ownership of NeoSpace

Research

A study of one firm's AI interviewer finds that a change of human reviewer hid a real improvement. Institute

Liu Zhang and Mark Esposito posted a preliminary working paper to arXiv in October on how to tell whether an AI agent in production is getting better or worse. They studied Zara, an AI agent that conducts live job interviews by voice and video for micro1, a firm that vets professionals for remote work. From March to August 2026 the firm's evaluation team scored 2,611 recorded interviews in 14 batches against a nine-part rubric, with reviewers overlapping across batches. One reviewer rated the same batches 0.79 standard deviations lower than the other. When the stricter reviewer joined, the raw quality trend went flat. Adjusted for the reviewer, quality rose by 0.53 standard deviations over the period. The authors also estimate how quickly evidence goes stale. At the rate of change their records cannot explain, uncertainty reaches the size of the largest effect of any logged release after about five weeks. The range is wide, from one to 21 weeks. Among the possible causes of that unexplained movement they list silent vendor updates under unchanged model names. Interview-related support tickets fell about 10% a month against technical tickets, which they read as agreement in direction, not proof that evaluation caused the gain. They say nothing in their model is specific to interviewing, and that the agent could as well answer support queries, execute transactions or write code. The evidence comes from one firm and one hiring agent, in a paper not yet peer reviewed. It still names a failure that a standard validation report would miss: a quality score that moves with who does the scoring rather than with the system. A bank that rates its agents with small, changing panels of reviewers is exposed to the same effect, and an agent built on a vendor's model can change under a validation report within weeks.

Liu Zhang and Mark Esposito (arXiv): Reliability of AI Agents: Rater Effects, Drift, and the Return to an Evaluation Program

Security

OWASP's lab tests find offensive AI agents still needed a person to finish, and back checks that sit outside the model. Institute

The OWASP GenAI Security Project, part of a nonprofit that publishes software-security standards, released a 57-page study in October comparing two ways of using AI to attack computer systems. In one, a person works step by step with a chat model. In the other, agent frameworks are set loose to work on their own. Over about fifteen months the testers aimed both at two Hack The Box machines, practice targets released after the models' training data was gathered. The agent track covered four frameworks in nine setups, among them Claude Code and OpenAI's Codex, running Anthropic's Opus 4.8, OpenAI's GPT 5.5 and an open model run locally. No agent setup completed an exploit on its own. The strongest multi-agent setups found a working route to taking over an account, then the providers' safety filters stopped them and a person ran the last step. In the other track, an operator working like a novice, with a chat window and basic tools, reached a simulated administrator takeover. Across both tracks, people mostly had to step in for ordinary problems such as diagnosing the environment, breaking loops and recognising success. The report cites other studies on two further points. The software wrapped around a model moves results far more than the choice of model. Benchmark scores overstate success against live targets by two to eight times. It recommends that high-impact or irreversible agent actions be cleared by a fixed authorisation check outside the model. A model's own safeguards are not enough, it says, because an agent may get past them by asking again.

The study tested only models on general release, not those held back for selected partners such as Anthropic's Mythos Preview. It ran each setup once, on targets nobody was defending, and tested no bank. It does not show that AI cyber risk is overstated. It gives a way to read capability claims: a score says little without the software around the model, the number of attempts and how often a person stepped in. Its control advice also applies to a bank's own agents, where a payment, a data export or a production change can be gated by a check the agent cannot talk its way past. In the authors' runs the providers' safety filters were the main barrier, and a model run locally has none.

OWASP GenAI Security Project: LLM Exploit Generation Phase 2

On the radar

• Dutch police arrested two men on 6 October over dozens of AI-made bank-helpdesk recordings that led about 170 victims to fraudsters, who took about EUR 380,000 through bank cards held to a phone. Dutch National Police