BankingNewsAI Daily Brief ·
OpenAI launches ChatGPT Wallet so AI agents can make purchases with Amex, Mastercard and Visa
Banking AI
Financial institutions & fintech technology
OpenAI introduces ChatGPT Wallet for agent purchases with Amex, Mastercard, Visa
For the head of card fraud at a regional issuer
OpenAI names seven fraud and payment firms it is helping recognize agent purchases, and no issuing bank besides American Express. Your fraud model meets these charges cold, and your dispute team meets an approval record it did not write.
ChatGPT Wallet lets an AI agent buy things with a customer's card without holding the card. Before each purchase the customer sees the items, total and card, then approves with Face ID or a fingerprint. The network issues a payment token, with its own number, security code and expiry, and the merchant charges that. Your bank sees an online charge on a real account. Simple model: a tokenized purchase with a strong customer check. The imperfection is that a bot did the shopping, and OpenAI says risk systems may flag these and decline them. The exception is disputes. Each approval creates a record of what the user authorized, and your dispute team has no rule yet for using it. Rollout starts with US ChatGPT Pro subscribers. Pre-approved spending limits come later, and then nobody touches the fingerprint reader at purchase time.
→ Action
Fraud and disputes: Flag ChatGPT Wallet tokenized transactions in authorization data, track their decline rate, and set evidence rules for OpenAI's per-purchase approval records.
Full story →Read article → from OpenAI
Fed survey finds AI risks are biggest threat to financial system
For the chief risk officer at a regional bank
Nearly nine in ten respondents now name AI as a top risk, up from half in the spring. Examiners will look for AI in your loan book and your cyber defenses, not your strategy deck.
The Fed asked market participants, economists and analysts what could break the financial system. Through Oct. 28, 87% named AI, up from 50% in the spring. The next item, a fall in risky assets, drew 52%. The AI worry has three parts. Stock prices may assume earnings that never arrive. State actors, or frontier models themselves, could cause cyber damage. And complex financing has been built to pay for AI. The simple model is one bet seen from three sides: if earnings disappoint, prices fall and that financing gets harder to repay. This is a survey of opinion, not a stress test. The exception is hedge fund leverage, the highest since 2013 and held by a small group, which can turn a price drop into forced selling. A bank's AI risk sits in its credit files.
→ Action
Credit risk: Tag loans and counterparty lines tied to AI-related financing so you can report total exposure when examiners ask.
Full story →Read article → from Moneywiredaily
Plaid powers personalized financial insights with Google Wallet in Gemini app
For the head of digital banking at a regional bank
Your customers can now ask Gemini which subscriptions went up this year, answered from your transaction data via Plaid. You hold the accounts; Google holds the conversation. Decide whether to support that feed or limit it.
Gemini users in the US can now link bank accounts, credit cards, loans and investment accounts to Google Wallet through Plaid, then ask things like which subscriptions rose in price this year. The chain is short: your bank holds the records, Plaid pulls them, and Gemini reads them next to the customer's Gmail, Photos and Search. Access is read-only. Gemini cannot move, manage or invest money, and customers can switch it off. So if no money moves, what is lost? The conversation. Budgeting advice is where customers decide where to save, borrow and open the next account. It now happens in Google's app while your bank keeps the ledger, like a shop that stocks the shelves while someone else talks to the customers. Read-only describes today.
→ Action
Open banking partnerships: Check whether your Plaid data access agreement covers AI assistant use, and whether you can see which connections feed Google Wallet.
Full story →Read article → from Plaid
General AI
Large language models & AI infrastructure
Export-banned Nvidia chips are behind China's most powerful AI models
For a model-risk officer at a regional bank
If DeepSeek V4 or Kimi K3 runs anywhere in your bank, the model file itself now carries evidence of training on chips the U.S. barred from China. The download was free; the provenance question is not.
An open-weight model is a file of trained numbers that anyone can download and run. Will DePue of Unemployment Labs read those numbers for traces of the hardware that made them. He says Moonshot's Kimi K3 and DeepSeek's V4 bear fingerprints of Nvidia Blackwell chips, which are banned from export to China. Why should a bank that never bought a chip care? The chip's marks travel with the file, like a serial number stamped on a part. The evidence is imperfect. The chips could have sat in an offshore data center, in Malaysia say, which may be legal. The exception is Moonshot, which a White House adviser, citing U.S. intelligence, said used banned chips. Know which systems run these files. The free model may carry the longest paper trail.
→ Action
Model risk: Flag DeepSeek V4, Kimi K3 and other Chinese open-weight models in the model inventory, including vendor products built on them.
Full story →Read article → from TechCrunch
Anthropic's potential record IPO draws orders for six times the shares
For the chief procurement officer at a regional bank
Your Claude contracts were signed with a private lab. Soon the counterparty answers to shareholders who paid for a $1.4 trillion valuation and expect revenue to keep climbing. Redo the concentration assessment before the listing, not after.
Anthropic, maker of Claude, is going public. It is raising up to $64.4 billion in the IPO plus a $26 billion private placement, at about $1.4 trillion fully diluted. Orders run about six times the shares on offer. The number selling it is an $81 billion revenue run rate, up from $65 billion at the end of July. That figure takes 28 days of usage fees, or one month of subscriptions, and multiplies them out to a year. It is like judging a shop's annual sales from one good month: fair while usage rises, flattering if it dips. One exception matters. Nvidia, which sells Anthropic tens of billions in chips, is buying about $10 billion of the shares. Why would a supplier's IPO matter to a customer? Your bank's spend is part of the number they bought.
→ Action
Third-party risk: Add Anthropic and its Nvidia chip dependency to concentration reviews, and check price-change notice terms in Claude contracts.
Full story →Read article → from Businesstimes
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime