Reading time ~15 mins

JPMorganChase asks leaders to time the decision on a software flaw and the fix separately, and India's cyber agency sets rules for auditors' own AI tools. Goldman takes AI coding tools back from most traders and bankers. Visa finds shopping agents often fail and buy anyway, and Plaid connects US bank accounts to Claude. The Bank of England puts AI's productivity lift so far at a fifth of firms' hopes, and a study measures EU banks' reliance on non-EU providers. Nate B. Jones reads US data-centre votes as a gain for the biggest suppliers.

Perspectives

Nate B. Jones reads the US votes against data centres as an opening for the largest AI suppliers Perspective

Nate B. Jones, an independent AI commentator, published an executive briefing on his paid Substack on 8 November. Its title asks whom a company will rent AI capacity from in 2027, and at what price, if the backlash against data centres leaves fewer suppliers. His subtitle says US voters went against data-centre development in 40 of 41 local ballot measures on 3 November, a count that was not checked here. Use keeps rising all the same. He cites a May survey by Edison Research at SSRS in which 65% of US adults had used an AI platform in the previous week, up from 52% in February. In an October poll by ABC News, the Washington Post and Ipsos, 72% opposed a new data centre in their community. Nate B. Jones writes: "People are finding AI useful in their own lives while rejecting the infrastructure being proposed down the road."

His argument starts from demand. If people keep using AI services and companies keep buying them, someone will supply the capacity, and where that happens decides who gets the investment, the work and the payments. Nate B. Jones writes: "Developers that can’t afford a delay will have to sell." Bigger operators, he says, can then buy projects cheaply and face less competition from new facilities. He expects the next two years to reward the companies and communities that can actually deliver capacity, so a backlash against the industry hands some of its most powerful firms an opening. He also blames the industry itself. Silicon Valley, he writes, had every chance to explain how AI gives people more room to learn, build and earn, yet people hear that it automates their work and makes its builders rich.

Readers have met the backlash from other sides. This brief carried Dutch grid, permit and consent limits on 16 September and a lender's checklist from the World Economic Forum and Oliver Wyman on 29 September. The cep think tank's paper on Europe's compute gap followed on 30 October. On 5 November three US banks arranged first-lien debt in a $63 billion vehicle that leases AI chips to Anthropic. His new point is that local refusals concentrate supply rather than reduce it. That rests on his reasoning, not on data, and the votes were local and American.

His free list of contents says the briefing works through what a one-year delay costs a hypothetical project carrying $500 million of debt. It also covers which contract terms decide whether a smaller provider survives the wait. It explains why slower construction makes overbuilding less likely, since demand gets time to catch up, and how to tell whether a failed developer ran out of customers or of cash. It covers what Canada, the EU, the UK and Singapore offer developers, and ends with a 90-day plan for finding which AI commitments depend on capacity that does not yet exist. For subscribers, his post also takes up whether a second supplier protects a buyer when capacity runs short.

A bank meets his argument on two desks. As a buyer of cloud and AI capacity for 2027, it may face fewer suppliers at renewal. As a lender to data-centre developers, it may see a borrower's project sold because the owner could not carry a delay, which makes a project's opening date part of the credit question.

Nate's Substack (By Nate B. Jones)

A security industry body says AI agents call for a reorganised security team, and for retraining juniors before cutting staff Perspective

Jakub Szarmach, who curates AI governance resources, shared a Cloud Security Alliance white paper on LinkedIn on 6 November, Designing the AI-First Security Organization. Rich Mogull is its lead author, and it draws on three summits of security chiefs held under the Chatham House Rule in the US in May and June 2026. The paper is undated, and the alliance, an industry association, binds no one. Its premise is that frontier AI models now do a real share of hands-on security work, such as finding software flaws and reviewing code. A team's capacity has come loose from its headcount, and queues and approval chains built for human speed become the constraint. Jakub Szarmach writes: "AI-first security does not mean adding agents to the existing org chart."

The paper sorts eleven security functions into seven kinds of change, each with its own pace. Finding and fixing software flaws merges into one continuous pipeline within months. The identities that software and agents use need attention now. Compliance and supplier risk become heavily automated, while the human sign-off narrows to attestation and risk acceptance, where personal liability sits. The wider restructuring takes one to two years.

Jakub Szarmach writes: "The governance model is particularly strong." Agents act inside pre-approved, monitored limits, and people supervise the pattern of outcomes rather than each action. A named person must approve any action that cannot be undone, could disrupt production, is visible outside the firm, could raise access rights or is attested to a regulator. That check has to sit outside the agent, because a manipulated agent can misjudge its own action. The paper does not say where the line falls in a given firm. It asks firms to draw it, document it and staff it before an incident forces the question.

The workforce section is the less comfortable part. Entry-level queues were the profession's apprenticeship, and agents are taking them. The paper asks firms to restructure before they resize, moving staff into seats where they check agent output under senior review. It warns that sign-offs turn cosmetic when people approve well-formatted output without checking it, and asks for two-person approval of changes to fleet rules and emergency stops. Its main measure is validated risk reduction per security employee, and it notes that cutting staff flatters that ratio without removing any risk.

The evidence is summit discussion, with no measured outcomes and no bank or EU content, and the alliance sells training and certification. Readers had Bank of America's 4,000 people fixing flaws found by frontier AI on 20 October and the alliance's AI Controls Matrix on 6 November. Today's JPMorganChase guidance asks how fast a bank fixes a flaw. This paper asks who in the security team does the fixing and the checking, as 2027 security budgets and graduate intake are set while agents make cuts look affordable.

Cloud Security Alliance via LinkedIn (By Jakub Szarmach)

Netherlands & Sovereignty

An outside measurement of 452 EU bank domains scored banks at 35.1% on independence from non-EU providers. Institute

Jürgen Brandl and co-authors, from TU Wien and other institutions, presented a measurement study at the ACM Internet Measurement Conference, held 12 to 16 October. They checked 1,988 domains of public institutions, banks, newspapers and universities across all 27 EU member states, 452 of them banks. They used only what is visible from outside, such as who runs a domain's name servers, hosting, certificates and email. Only 61 domains, about 3%, showed no dependency on a non-EU provider. A typical domain relied on non-EU providers in four of the six layers, with Microsoft, Google and Cloudflare dominating. On the paper's equal-weight score, banks reached 35.1%, against 51.3% for government and 51.1% for universities. Microsoft routed email for 38.9% of the bank domains. The paper names no bank and gives no Dutch breakdown. The authors call their score a jurisdictional proxy, not a resilience measure, and favour tested ability to switch providers over rules requiring local ones. Readers had the policy arguments on European dependence from cep and ECIPE in October. This is a measured figure for banks, and the method needs no access to a bank's systems, so anyone can run it on a bank's own domains.

Association for Computing Machinery: EU Digital Sovereignty in Practice

Industry & competition

Goldman Sachs has taken AI coding tools away from most traders and bankers after months of use. Media

Bloomberg reported on 2 November that Goldman Sachs restricted most traders and bankers from advanced AI coding tools, among them Anthropic's Claude Code and Microsoft's GitHub assistant, under internal guidance updated in October. These tools write software from plain-language instructions. Staff had used them for several months to monitor client trading flows, model possible market and position changes and produce written and visual material. Access is now limited to engineers and to "strats", staff who combine trading, quantitative research and software work. A Goldman spokesperson said a period of experimentation had assessed productivity and risk management. The bank puts tools "in the hands of those who are best suited to leverage them for the long term", he said. In July its chief information officer, Marco Argenti, had praised traders and bankers building prototypes this way for the firm's roughly 12,000 engineers to refine. Bloomberg's account gives no findings from the review. The decision draws a line many banks have left open, between business staff who build working software with these tools and the engineers who run it in production. One large bank drew it after trying the wider version.

Bloomberg News

Innovation

Visa's first benchmark of AI shopping agents found they failed 42.8% of runs, and most failures still ended in a purchase. Vendor

Visa launched its Agentic Commerce Benchmark on 5 November. The first round tested 22 AI models from 12 labs on 15 shopping tasks, such as buying a television, a laptop or a personal-care appliance, with three attempts at each. Overall, 57.2% of runs passed. The best model passed 82.2% of its attempts and the worst 28.9%. All three attempts succeeded in only 34.2% of model and task pairs. In 54% of failed runs the agent still bought something, either a product that missed the customer's stated needs or one chosen before it had found them all. Visa does not name the models, and it designed and scored the test in a market where it has a commercial stake. The purchase-despite-failure figure is the one that reaches a card issuer. A purchase the agent completed against the customer's wishes is not fraud and not a merchant error, so it may fit none of today's dispute reasons cleanly. Visa's own test says such purchases are common.

Visa

Plaid now connects US users' bank accounts, read-only, to a personal finance feature in Anthropic's Claude. Vendor

Plaid, which links bank accounts to financial apps, announced on 5 November that it powers account connections for Personal Finance, a new feature in Anthropic's Claude assistant for US users. Claude's access is read-only, so it cannot move money. Plaid says Claude can track spending across accounts, spot unexpected fees, plan debt repayment and explain what an interest rate costs. It says Anthropic joins more than 800 AI companies building with Plaid. Nearly two-thirds of people who connect bank accounts to AI assistants through Plaid had never used a Plaid-powered finance app before, it says. These are Plaid's own figures. The feature is US-only. A general-purpose assistant now answers a customer's questions about their money from the bank's own transaction data, without the bank in the conversation. In the EU the comparable route is a licensed account-information service acting on the customer's consent under the payment services rules.

Plaid

Research

The Bank of England puts AI's lift to UK productivity so far at 0.1 to 0.2 points a year, against firms' hopes of 0.9. Authority

The Bank of England's November Monetary Policy Report, published 6 November, cites its Decision Maker Panel, a regular survey of UK firms. Firms estimate that AI raised their productivity by around 0.5% over the past three years. The Bank says that points to AI adding about 0.1 to 0.2 percentage points a year to productivity growth since early 2024. Firms expect about 0.9 points a year over the next three years, split roughly equally between slower employment growth and faster output. So far they report close to zero effect on employment overall, which the Bank says matches empirical studies. It attributes most recent productivity strength to cyclical factors and sees growing upside risks from AI. All of these are firms' own estimates, not measured output. The report gives a central bank's number for the gap between what AI has delivered and what firms expect, roughly a factor of five. It also places half of the expected gain in slower hiring, which bears on workforce plans that count on AI savings.

Bank of England: Monetary Policy Report, November 2026

A Bruegel economist argues that oversight of AI agents fails if agents can alter the records or monitor copies of themselves. Institute

Bertin Martens, an economist at the Brussels think tank Bruegel, published Why AI goes rogue and what can be done about it on 27 October. He groups the risks of many agents working together into four: collusion, divided responsibility, correlated errors, and speed or scale that overwhelms human review. Rules and technical safeguards alone, he argues, cannot contain agents that misbehave. He draws on the economics of contracts and crime. Random audits deter only if evidence survives concealment, agents cannot know in advance what will be checked, and sanctions are severe enough. Agents able to change records may answer tougher audits by hiding violations. Diverse agents with linked incentives can watch one another, while swarms of identical models let shared errors reinforce each other. The piece is a synthesis of research, not a tested control design. Its two conditions can be checked in any bank's agent design: whether an agent can write to the logs an audit relies on, and whether the monitoring model is the same one it monitors.

Bruegel: Why AI goes rogue and what can be done about it

A study for the EU AI Office takes bank governance as the model for who should own risk at frontier AI labs. Institute

The EU Publications Office released Risk responsibility allocation in the EU GPAI Code of Practice on 28 October. Researchers from the Centre for Long-Term Resilience, SaferAI and GovAI wrote it for the European Commission's AI Office, and it reflects only their views. Peter Slattery, an MIT researcher, shared it on LinkedIn on 5 November without comment of his own. It reads Commitment 8 of the voluntary code for general-purpose AI models, on who inside a provider owns systemic risk, as the three lines model banks use. It cites EBA governance guidelines, the Capital Requirements Directive and BCBS 239 as the reference. The executives who run research and product should own the risk, with a risk officer challenging them, rather than the chief executive. An oversight body should be able to veto or reverse a deployment, and any override should go to the full board with recorded reasons. An EU subsidiary's board should not hold oversight when testing and deployment are decided at the parent, as that oversight would be nominal. Nothing in it binds a bank that only uses these models. Its yardstick for an AI provider's governance is the one bank supervisors already apply to banks.

European Commission via LinkedIn (shared by Peter Slattery)

The OECD finds one in five firms used AI in 2025, against half of large firms, and expects the gap to widen. Institute

The OECD published its Digital Economy and Society Outlook 2026 on 5 November. In 2025, 20% of firms with at least 10 employees used AI, against 52% of firms with 250 or more. The OECD's projections show that gap widening through 2030, while the gap in cloud use narrows. It names skills, data readiness and organisational change as the barriers. Among adults, about 37% used generative AI in 2025, 19% of them at work. For a bank whose business clients are mostly small firms, the OECD expects the AI gap between those clients and large corporates to grow, not close.

OECD: Digital Economy and Society Outlook 2026

Security

JPMorganChase tells leaders to measure two clocks on software flaws: time to decide and time to fix. Corporate

JPMorganChase's global technology leadership team published 200 days into the frontier: What we know and what to do now on its technology blog on 2 November. It says AI is speeding up the discovery of software flaws and the automation of attacks, while repair is not keeping pace. Its sharpest point is about time. A firm may decide in hours and still need days or weeks to act, because it must find the affected systems and their owners, test a fix, coordinate suppliers, deploy and confirm. So the bank asks leaders to measure decision speed and execution speed separately. It treats technology whose maker no longer issues security fixes as a time-sensitive business liability. Every accepted exception needs an accountable owner, funding, compensating controls and a dated exit plan. Suppliers can set the pace of any fix, so it asks for agreed notification and escalation terms with critical suppliers and for exercises that test executives' decisions, not only patching. Its advice on AI agents repeats what readers had from the FSB, Hong Kong and the OCC. These are recommendations, with no evidence of how far the bank applies them. Readers had Bank of America's 4,000 people fixing flaws found by frontier AI on 20 October, and the ECB's Pedro Machado calling fast patching banks' biggest challenge on 2 November. The largest US bank has now written down how it measures that problem.

JPMorganChase: 200 days into the frontier

India's cyber agency sets out how auditors should check AI systems, including the AI tools the auditors use themselves. Independent

CERT-In, India's national computer emergency response team, issued AI Security Auditing Guidelines v1.0 on 4 November. Jakub Szarmach, who curates AI governance resources, shared the 107-page document on LinkedIn on 5 November. Jakub Szarmach writes: "AI governance needs evidence, not just policies." Four risk tiers, from Low to Critical, set how intensive and independent an audit must be across nine audit areas. The part he singles out is AI-assisted auditing. Tools must be validated on independently reviewed cases before auditors rely on them. AI output stays apart from original evidence, and auditors do not hand scope, severity, opinion or closure to the tool. Audit tools are tested against documents that try to change the audit criteria, and material AI use is disclosed in the report. Jakub Szarmach writes: "An audit report is not assurance unless its conclusions can withstand scrutiny." For AI agents, it lists tests such as bypassing approval steps and checking that the emergency stop works, and it requires retests after the provider changes its model version. The text is addressed to organisations running AI in India or handling Indian citizens' data, and names no legal basis, start date or penalty. It creates no duty for a bank in the EU. Where internal auditors draft findings with AI tools, this public method treats the auditors' own AI as something to control.

CERT-In via LinkedIn (shared by Jakub Szarmach)

On the radar

• Microsoft reported on 6 November that a Cambodia-based criminal group it calls Storm-3358 began using ARTEX, an AI framework that automates intrusions, against financial firms in October, mainly in Southeast Asia. Microsoft Security Blog