BankingNewsAI Daily Brief ·
Google releases Gemini 4 Carbon first to cyber defenders
Banking AI
Financial institutions & fintech technology
Bank of America asks customers not to share passwords with AI assistants
For the retail digital-banking head at a regional bank
BofA has told customers in writing that whatever an agent does with a shared login is their problem. Your account agreement probably says the same; your dispute team will still be the one taking the call.
Bank of America emailed customers asking them never to give their username and password to an outside service, "including an AI-enabled tool or agent that offers to act on your behalf." The risks run in a short chain. The agent moves money unexpectedly, the bank reads the login as an intruder and locks the account, or a flaw in the tool leaks the password to hackers. The email adds that anyone who shares is responsible for what follows, as most account agreements say. Handing over a password is lending a stranger your house key. Except here the customer picked the stranger to shop for them. Why not issue agents their own key? Banks and processors have not settled how. Until then, the contract says consent and the fraud screen says theft.
→ Action
Legal and compliance: Check that account agreements and unauthorized-transaction dispute procedures address activity by AI agents using customer-shared credentials.
Full story →Read article → from TechCrunch
PNC CEO says AI will drive more bank mergers, especially small banks
For the ceo of a community bank
Demchak expects banks still running mainframes to become targets within a few years, while PNC spends hundreds of millions to save billions annually. Your AI budget is now your independence plan.
Demchak's bet: AI pays off for a bank that owns its compute (its own data centers and graphics processing units) and its own data. PNC will spend hundreds of millions to save billions annually. Banks still running mainframes, or short of investment dollars, cannot make that trade, so he expects a lot of mergers, particularly among small banks. Why isn't PNC buying? Folding a large deal onto one platform takes two or three years, and he says the game could be over in two. The gap: PNC still buys vendor pieces that fit, and turned on a third-party mortgage interface in a weekend. The question for a smaller bank is which layer it owns. Renting furniture is fine. Rent the foundation and someone else decides when you move.
→ Action
Technology strategy: List which AI workloads and data sit on owned versus vendor systems, and check exit terms and lock-in clauses in each contract.
Full story →Read article → from Banking Dive
Adclear AI cuts NatWest marketing approval time by 80%
For the chief compliance officer overseeing marketing review
NatWest cut approval time 80% by putting AI checks inside the workflow its marketers already use. Speed is the easy part to copy; the record of who approved what, and why, is what your examiner will ask for.
Adclear is software that reads a bank's ads, websites and social posts before publication and flags copy, creative or disclosures that break the rules. NatWest wired it into Workfront, its marketing workflow system. That replaced separate review routines for current accounts, cards, loans, mortgages and savings. Approval time fell 80%. Manual review delays launches three to five days; cutting that 80% turns a week of waiting into a day. The simple model: software does the first read, and people sign off on what it flags. The gap: NatWest reports speed, not error rates. Is faster safer? Not alone; FCA interventions on promotions doubled last year. The exception is the audit trail: one record instead of five routines. For a US bank, that record, not the 80%, is what examiners will test.
→ Action
Marketing compliance: Pull 20 recently approved ads and confirm each file shows reviewer, rule checked, disclosure changes and sign-off date.
Full story →Read article → from Regtechanalyst
General AI
Large language models & AI infrastructure
Google introduces Gemini 4 Carbon, releasing it first to cyber defenders
For the CISO of a regional US bank
Carbon scores 88.2% at finding flaws and 71.0% at patching them, and the same model scores 54.8% at exploiting them. Defenders get it first; your patch calendar should assume attackers get it next.
Gemini 4 Carbon is Google's newest AI model, and Google's teams have seen it work on code for days at a stretch. On security benchmarks, run without guardrails, it scores 88.2% at finding vulnerabilities, 54.8% at exploiting them and 71.0% at writing patches. Why build an exploit tool at all? Because finding, proving and fixing a flaw are the same skill. Google's only lever is order. A subset of cyber defenders gets it this week through the Fairwind Program. Developers and businesses follow "as soon as possible" at $2 per million input tokens. Benchmarks are not your codebase. But Argon shipped only two weeks earlier. A locksmith's pick opens every door, whoever holds it. Assume the gap between your scan and an attacker's is weeks.
→ Action
Vulnerability management: Shorten critical-patch deadlines for internet-facing systems to match faster AI-found exploits.
Full story →Read article → from Google
ServiceNow and OpenAI expand partnership, launch ServiceNow plugin for ChatGPT
For the cio of a bank running servicenow
ServiceNow permissions were set for staff clicking through forms. Now a ChatGPT conversation can assess change risk and update the record. Decide whether those approvals still hold before a business unit switches the plugin on.
The new plugin lets someone working in ChatGPT reach into ServiceNow and act: find similar incidents, assess change risk, update records. ChatGPT does the reasoning. ServiceNow, through Action Fabric, applies its own business rules and permissions and executes the action. The simple model is that nothing changes, because the agent inherits whatever the user may already do. That model is imperfect. A permission written for a person who reads each form before saving becomes a permission for software that can save many forms in one reply. The exception is any step your workflow still routes to a named human approver. Those hold. A planned voice agent will even assign risk ratings to AI use cases. The software is filling in its own permission slip.
→ Action
IT access management: Audit which ServiceNow roles can update incident and change records, and require human approval for ChatGPT-initiated changes.
Full story →Read article → from Servicenow
Microsoft makes its Copilot harness available through Microsoft Foundry
For the head of model risk at a US bank
Your inventory approves one model per use case. Microsoft now offers to pick among 20+ models per task and keep changing the harness around them. Validation has to attach to the router, not the model.
A harness is the software wrapped around an AI model: the instructions, context, tools, sandbox and controls that turn a model into an agent. Microsoft runs its own Copilot products on one and now offers it through Foundry. Microsoft runs the agent loop and sandbox; the customer brings instructions, tools and model choice. The harness supports 20+ frontier models plus bring your own key, and can route each task to the right model or models. The simple model: the bank picks a model, validates it, ships. If the router chooses, which model did you validate? In practice, the routing policy, and a harness Microsoft keeps changing. Customers can still choose the model themselves. It is like licensing a driver, then finding the car swaps drivers at each light.
→ Action
Model risk: Require Foundry agent use cases to log which model handled each task and pin approved models before production.
Full story →Read article → from Microsoft
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime