BankingNewsAI Daily Brief ·
Nvidia tops earnings estimates as its board rules out Jensen Huang's return as CEO
Banking AI
Financial institutions & fintech technology
BNY and Microsoft complete first agentic negotiation of a US Treasury trade
For the head of markets at a US dealer bank
BNY wrote the rulebook for agent-to-agent Treasury negotiation, filed for a patent, and runs the ledger that checks both sides agree. Join its protocol or build your own audit record before a counterparty's agent asks which you use.
ANP is a control layer BNY built so software agents from two firms can haggle over a trade. Microsoft Treasury gave its agent instructions, and the agent reached BNY through a secure gateway. BNY checked the counterparty, security, size, settlement terms and pricing. Then the agents traded offers inside set limits. Each side sent the agreed terms separately to ANP's central ledger, which confirmed they matched. Only then did human traders on both sides approve, and existing systems booked the trade. If people still sign off, what did the agents save? The back-and-forth messages, and any later argument over what was agreed. The limit: this was one Treasury bill between two firms. Repo could come next. And the record that settles disputes sits at BNY.
→ Action
Markets compliance: Map how trader approvals of agent-negotiated terms would be recorded and reconstructed under your current trade surveillance and recordkeeping rules.
Full story →Read article → from Bny
Visa report finds AI-enabled fraud and third-party attacks scaling worldwide
For the head of third-party risk at a regional bank
Third-party agents sat behind about 56% of compromised payment credentials, and incidents involving them rose 62% in six months. Your vendor reviews probably still weigh each processor by contract size, not by how many doors its one key opens.
Visa's fraud team studied attacks from January to June 2026. AI is not inventing new frauds. It runs old ones faster. Picture an office building where the cleaning contractor holds one key to every floor. Third-party agents were linked to about 56% of compromised payment credentials, and incidents involving them rose 62% in six months. Steal the contractor's key and every office opens. The picture is imperfect, because defense scales too: card-number guessing attacks rose 35%, yet Visa blocked 200% more of them and related losses fell 21.4%. That win belongs to the network, which sees across banks. For one bank, the exposure sits in its vendors' systems. So why tighten vendor checks if Visa catches the guessing? Because a stolen key does not need to guess.
→ Action
Third-party risk: List processors and fintech vendors that store or touch card credentials; require breach-notice terms and access logs from each.
Full story →Read article → from Visa
JPMorgan Chase explores AI agents for autonomous financial transactions with strict controls
For the treasury and payments innovation lead at a regional bank
JPMorgan says building the agents was the easy part, and deciding who answers for them was the hard part. By the first half of 2027, your corporate clients will ask you to answer it.
Kinexys, JPMorgan's blockchain unit, ran a test with Hyundai Card. Hyundai's AI agent started a simulated foreign-exchange trade, with no person involved, to keep euros in its blockchain deposit account. The control is simple: draw a box around each agent. An agent cleared for $5 million cannot send a payment above that, because the system rejects it. That works for one agent following orders. It is less clear when two agents negotiate terms with each other, and that is the phase targeted for the first half of 2027. Who agreed to the price? The box limits the amount, not the judgment. JPMorgan has announced no commercial deployment. Kinexys already averages about $7 billion a day. A box caps the amount. It does not say whose name is on the deal.
→ Action
Corporate banking legal: Draft per-agent dollar limits, kill-switch rights and liability terms for client agent-initiated FX payments before clients ask.
Full story →Read article → from Indiatimes
General AI
Large language models & AI infrastructure
OpenAI launches GPT-6.2 Astra and Sol to negative reception from power users
For a technology chief running model procurement at a bank
Astra now costs the same as Anthropic's Fable after a 20% cut, while one tester watched it spend seven days producing code that ran 56% slower; a cheaper model you cannot steer saves nothing until your own tests agree.
GPT-6.2 Astra and Sol are OpenAI's new hosted models, sold by the token through ChatGPT and the API. After a 20% cut, Astra costs $8 per million input tokens and $40 per million output, the same as Anthropic's Fable. Sol costs a quarter of that. So why would a cheaper model with high benchmark scores be a worse buy? Because you pay per token, not per finished job. Matt Shumer gave Astra sub-agent work, and the Claude models supervising it redid it. Theo Browne saw it spend seven days and produce a compiler 56% slower. A cheap plumber who leaves the pipe half-fitted gets paid twice. These are a few testers on their own tasks, and Browne says Astra caught code-review bugs other models missed. Benchmarks grade the model. Your invoice grades the job.
→ Action
Model risk: Pin production API calls to current model versions and rerun bank workload evaluations on gpt-6.2-astra before any switch.
Full story →Read article → from Venturebeat
Microsoft challenges OpenAI, Anthropic with cheaper, specialized AI models
For the CIO managing AI spend at a regional bank
Microsoft says MAI Code 2 delivers 96% of Claude Opus 5.5's performance for 5% of the cost. That makes the budget case for narrow models. It also adds work for model risk: each cheap model needs its own validation.
Microsoft built models that each do one job, such as writing code or checking security, instead of one model that does everything. The pitch is price. Microsoft says MAI Code 2 matches 96% of Claude Opus 5.5's performance at 5% of the cost. That is like paying $1 for a $20 tool that works almost as well. MAI Cyber 2 Flash beats GPT-6 Astra on one security benchmark at under 2% of the cost per task. The simple model is narrow tasks on cheap models. It has limits. Microsoft picked the benchmarks, some prices are promotional discounts, and the coding model is weaker at advanced math. The exception is GitHub Copilot, which will carry MAI Code 2. Your developers may use it before anyone validates it. Cheap tokens still need an owner.
→ Action
Model risk: Inventory which models GitHub Copilot uses for bank developers and require separate validation before MAI Code 2 touches production code.
Full story →Read article → from Axios
Nvidia tops estimates as board rules out Jensen Huang's return as CEO
For the chief credit officer with data-center loan exposure
The chips behind your data-center loans just posted revenue up 103% from a year ago. The founder who set their roadmap resigned, and the board will not take him back as CEO. Underwrite the guidance. Watchlist the governance.
Nvidia sells the chips hyperscalers buy for AI data centers. Q3 revenue was $115.8 billion, roughly double a year ago. Q4 guidance is $130.0 billion, plus or minus 3%. Simple model: demand this strong keeps GPUs worth something and keeps tenants paying rent. The flaw: a quarter of sales says little about what a used chip fetches later, and gross margin is guided down from 74.8% to 71-72%. The exception is leadership. Huang resigned, the board ruled out his return as CEO, and the CFO is also interim CEO while a search firm looks. Why are shares still below pre-scandal levels after a beat? Buyers price the roadmap, not the quarter. Lenders should too.
→ Action
Credit risk: Flag GPU-backed data-center loans for watchlist review tied to Nvidia's CEO search, and stress collateral values for a slower chip roadmap.
Full story →Read article → from Crowdfundinsider
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime