BankingNewsAI Daily Brief ·
Global stock selloff spreads after the worst US market rout since 2020
Banking AI
Financial institutions & fintech technology
Plaid powers account connections for Claude's new Personal Finance feature
For the head of consumer digital at a retail bank
Your customers' balances will now power debt-paydown advice in Claude, read-only, with your logo nowhere on the screen. Either build that coach in your own app, or learn what your Plaid terms say about AI assistants first.
Claude, Anthropic's AI assistant, now has a Personal Finance tab. A user asks a question, Claude prompts them to log in to their bank through Plaid, and Plaid sends over organized, refreshed transaction data. Claude then answers with real numbers: spending across accounts, debt-paydown plans, what an APR costs, unexpected fees. Your fees, in many cases. Access is read-only, so Claude cannot move money. Then what does a bank lose if no money leaves? The conversation. Picture a stranger reading your statements at the kitchen table and telling you which card to pay first. Nearly two-thirds of people linking accounts to AI assistants through Plaid had never used a Plaid-powered finance app. Plaid serves more than 800 AI companies. The pipe your bank opened for fintech apps now feeds advisers.
→ Action
Legal and compliance: Pull your Plaid data-access agreement and confirm what it permits when the data recipient is an AI assistant like Claude.
Full story →Read article → from Plaid
OpenAI forms JPMorganChase partnership ahead of possible 2027 IPO
For the CISO of a regional US bank
Daybreak is open to approved organizations, not just JPMorganChase. So a regional bank can apply, or it can tell its board why JPMorganChase's security team gets OpenAI's cyber models and its own does not.
Daybreak is an OpenAI program that gives approved individuals and organizations access to its cybersecurity models, plus security tools and support. JPMorganChase gets those models as part of a deal for OpenAI to supply the bank with AI services. The simple model: the bank's security team points OpenAI's tools at its own systems and fixes what turns up. The source does not say what the models find, what they cost or how approval works, and Brockman says the two firms will "explore" the work. Why would a vendor give its best defenses to one bank? It is raising money, with about $11 billion committed against a $30 billion target. Approval works like a building pass: you have to ask for one. A bank that never applies has also made a decision. It just hasn't written it down.
→ Action
Information security: Ask OpenAI for Daybreak approval criteria and terms, and brief the board on whether the bank will apply.
Full story →Read article → from PYMNTS
Visa benchmark finds AI shopping agents fail almost half of test runs
For the head of card disputes at a regional bank
AI agents in Visa's test failed 42.8% of runs, and 54% of those failures still bought something. Your cardholder authorized the agent, not the TV. Decide now which side of that line your chargeback rules put the loss.
Visa built a test where an AI agent must talk to a shopper who half-knows what they want, find the real requirements, then buy the right TV or laptop, or nothing. The test put 22 models from 12 labs through 15 tasks, and they failed 42.8% of runs. In 54% of failures the agent bought anyway. Multiply: nearly one run in four ends in a purchase that misses what the customer needed. The best model passed 82.2%, but all three tries passed only 34.2% of the time, so a good agent is not a steady one. This is a lab test, and it stops before checkout and payment; Visa plans to add those. A chargeback asks whether the customer authorized the payment. Here the answer is yes, and the customer still got the wrong TV.
→ Action
Card disputes: Map agent-initiated purchases that miss the customer's stated needs to a reason code, and set who bears the loss: issuer, merchant or cardholder.
Full story →Read article → from Visa
General AI
Large language models & AI infrastructure
OpenAI launches GPT-6 Cyber at higher API price than Anthropic, Google
For the CISO of a regional US bank
OpenAI's GPT-6 Cyber wrote working exploits for 948 Linux kernel flaws and sits behind an application form at $100 per million output tokens. The form slows your bank. Whether it slows attackers for long is the bet.
GPT-6 Cyber is a security model. Point it at software and it hunts for flaws and writes exploits. OpenAI rates its cyber capability Critical, up from High for GPT-5.6 Cyber. In one Linux setup it built working exploits for 948 kernel vulnerabilities; GPT-5.6 Sol managed 166. Access runs through Daybreak Red: you upload your defensive workflows, OpenAI reviews them, and if approved you pay $20 per million input tokens and $100 per million output. That is twice Anthropic's Mythos 5.1 and five times Google's Argon Cyber Permissive. The simple model is a locked toolshed. It is imperfect: the benchmarks are OpenAI's own, and 948 exploits are not 948 new flaws. The exception is time. A lock matters only until the same tools ship without one.
→ Action
Security operations: Submit a Daybreak Red application naming defensive workflows, and budget token spend at $20 input and $100 output per million tokens.
Full story →Read article → from Venturebeat
Stock selloff goes global after worst US rout since 2020
For the chief credit officer at a large bank
Your data-centre and chip loans rest on an AI buildout that markets now think a Democrat-controlled Congress could curb. Futures point to a bounce, which is the easiest moment to skip re-testing the collateral.
Stocks fell across Asia and Europe after the worst US market rout since 2020. The worry is policy: investors fear a Democrat-controlled Congress could curb AI development. Chip hubs fell steeply. Taiwan posted record losses, and South Korea halted trading after an 8% drop. Investors also question whether Anthropic reaches its $1.5 trillion IPO valuation target or delays its debut. The simple model: AI-linked collateral is priced off share values and IPO hopes, so when those fall, coverage thins. That model overstates one day's move, and US futures already pointed to a bounce. So why re-test anything? Because a new law is more like a zoning change than a storm. Prices recover from storms. Zoning changes what the land is worth.
→ Action
Credit risk: Re-run concentration and collateral stress tests on data-centre, chip and AI-lab exposures, assuming new limits on AI development.
Full story →Read article → from Semafor
Salesforce's Agentforce Coworker usage grows 21x in three months
For the coo of a bank running salesforce
Salesforce's agent costs $50 a seat and inherits each employee's permissions, which at most banks were written for people who log off at six. Write the logging and model-risk rules before business lines buy seats.
Agentforce Coworker is an assistant inside Salesforce. An employee hands it a job, such as checking an at-risk renewal. It plans the steps, calls other agents and workflows, and returns results for a person to approve. It works from the records, permissions and sharing rules a bank already has. Usage grew 21x in three months, to more than 185,000 users, at $50 per user per month. The simple model is that the agent can do only what its employee could. The flaw is that approval and audit cover "supported operations," not all of them. New Microsoft 365 and Google Workspace links also pull email and files in. Scheduled tasks start in December, so jobs will run while nobody watches. A permission set written for a person will then govern something that never goes home.
→ Action
Model risk: List which Agentforce Coworker actions lack approval and audit coverage before any business line turns on seats.
Full story →Read article → from Salesforce
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime