BankingNewsAI Daily Brief  · 

Lean community confirms Levent Alpöge's proof of the Riemann hypothesis is correct

🏦 2 Banking AI🤖 3 General AI

Banking AI

Financial institutions & fintech technology

2 stories

OakNorth pauses software business expansion to rebuild around AI

For the third-party risk head at an OakNorth client bank

OakNorth says bank governance reviews slow it down, so it paused expanding its software business for banks. Its new US charter makes it a lender competing with community and regional banks. Your vendor now sits across the table.

OakNorth is a London digital bank that also licenses its credit-analysis tools to lenders such as PNC, Capital One and Fifth Third. It has paused expanding that software business to rebuild its own systems around AI. Its CEO says taking tools through client banks' governance would slow it down. Its new US charter adds accounts, cards, payments and foreign exchange to the $3.5 billion it has already lent here. Its targets are commercial real estate and companies worth $5 million to $100 million, and its named competitors are community and regional banks. The pause covers expansion, not existing contracts, and the software may come back repackaged. Still, the newest version of the tool now serves the competitor first. You are renting the kitchen of the restaurant across the street.

→ Action

Third-party risk: Pull OakNorth contracts and confirm roadmap commitments, support levels, data-use limits and exit terms before the next renewal.

Full story →Read article →  from Finainews

Adyen taps Klarna for agentic commerce payment options beyond cards

For the head of card products at a regional bank

An AI agent checking out through Adyen used to have one kind of option: a card. Klarna is now the second, turned on with no integration work. Your card now competes for a checkout nobody is standing at.

Adyen Agentic is Adyen's checkout kit for purchases an AI agent makes for a shopper. Cards were its only payment type. Klarna is now the first other one. A merchant already taking Klarna through Adyen can switch it on with no integration work. The shopper sees the usual choices (pay in full, interest-free installments or financing) and confirms before the order completes. The simple model: the agent offers what is already wired in, so being wired in is the contest. That model is imperfect, because a human still approves each purchase. The exception is that Klarna published its handler as an open specification. Why give it away? So every payment provider can wire it in, and Adyen is first. For your bank, the card is no longer the only key on the agent's ring.

→ Action

Card products: Ask your card network and Adyen-using merchant clients how your cards are shown and defaulted in AI-agent checkouts next to Klarna.

Full story →Read article →  from PYMNTS

General AI

Large language models & AI infrastructure

3 stories

Lean community confirms Levent Alpöge's proof of the Riemann hypothesis is correct

For the chief model risk officer at a large bank

This proof runs 17 million lines, more than any person reads, and dozens of independent machine checks settled it. Your validation still trusts a reviewer reading the reasoning, which works until the reasoning outgrows the reviewer.

Lean is software that checks mathematical proofs for errors. Stanford mathematician Levent Alpöge wrote his Riemann hypothesis proof as 17 million lines of Lean code. Dozens of people ran it independently and got the same result. The checkers are built so they provably cannot accept a flawed proof. Trust moved from the author to the checker, the way a bank trusts the till count rather than the cashier's account of the day. The model is imperfect. A checker confirms the proof matches the statement as written, not that the statement says what people meant. Metaculus forecasters expected this in 2035, and Terence Tao said 2050. Validation that depends on someone reading the reasoning will meet outputs too long to read sooner than planned.

→ Action

Model risk: List high-stakes model outputs where validation depends on human review of reasoning, and flag which an independent automated check could confirm.

Full story →Read article →  from The-decoder

US tech stocks post worst week since pandemic on AI doubts

For the chief credit officer at a regional bank

Nvidia fell 14% on a reported SEC accounting review, and the Nasdaq 100 now sits 20% below its October peak. Prices moved in a week. Your covenants test quarterly.

This is a stock sell-off, not a default wave. The Nasdaq 100 fell 10.4% in a week, its worst since Covid, and sits 20% below its October peak. Nvidia dropped 14% on reports the SEC is reviewing its accounting; Micron, Broadcom and Sandisk fell too. Anthropic slid 9.1% a day after listing. The simple model: lower prices mean thinner collateral and harder refinancing for AI builders. It is imperfect: a share price is not a cash flow, and a data center runs no slower in a bad week. The exception is accounting. A price is an opinion; an SEC review can change the numbers the opinion rests on. High rates, held up by oil costs from the Iran war, remove the cheap rescue. The worry is not the drop. It is whether the books were true.

→ Action

Credit risk: Pull exposures to AI-infrastructure and chip borrowers, and rerun collateral values on loans secured by their shares or GPUs.

Full story →Read article →  from The-decoder

OpenAI requires approval for gateways and routers reselling its models

For the head of third-party risk at a regional bank

Your vendor inventory probably lists the AI tool, not the router behind it. OpenAI now decides whether that router keeps access, so the contract that matters most may be one your bank never signed.

A router such as OpenRouter is one account that reaches hundreds of AI models from the US, Europe and China, instead of a contract with each lab. OpenAI now requires its review and approval before any gateway or router resells its models. The worry is distillation: Anthropic and US security agencies say Chinese labs copied American models through resellers that hide the end user. Those cases mostly involved gray-market "transfer stations." The rule covers every intermediary anyway. So why would a bank's legitimate router be at risk? Probably it isn't. But approval is no longer automatic. Like a spare office key held by a cleaning contractor, it works until the landlord changes the locks. Your bank's AI access may rest on a permission it never asked for.

→ Action

Third-party risk: List every vendor and internal tool reaching OpenAI models through a router, and confirm that router has OpenAI approval.

Full story →Read article →  from Semafor

Get this in your inbox every morning

Free · No spam · Unsubscribe anytime