BankingNewsAI Daily Brief ·
Microsoft warns that AI agents pose insider threats and that frontier models cannot be trusted
Banking AI
Financial institutions & fintech technology
Huntington launches 1866 Studio, an AI small-business operating platform
For the head of business banking at a regional bank
Huntington is offering an AI back office to small businesses that bank elsewhere, betting the relationship follows. Your rate sheet now competes with a tool that writes their ad copy. Decide build, buy or partner before neighbors copy it.
1866 Studio is a set of eight AI agents, built on Google Gemini, that do the chores a small-business owner hates: licensing a new company, drafting ad copy and social campaigns, and spotting what will trip up a financing request. The base level is open to any small business, customer or not. Why hand it to people who bank elsewhere? Because Huntington, with about 500,000 small-business customers, treats it as acquisition. The owner opens the tool every day, and the account tends to follow, the way people shop where they already park. The model is imperfect: ad copy is easy to get elsewhere. The exception is the financing check, which sits right beside a loan. Banks competing on deposit rates are now pricing against a habit.
→ Action
Business banking product: Price out building, buying or partnering for an AI tool set covering licensing, marketing and loan-readiness for small-business prospects.
Full story →Read article → from Banking Dive
Santander turns to Nvidia to build new AI banking capabilities
For the CIO at a mid-size US bank
Santander says some tasks fell from €5 to about €0.03 per interaction, roughly 170 times cheaper. Its one detailed customer-service case saved more than 50%. Find out which of those two numbers your AI contracts are priced against.
Santander built a test bench called Model Arena. It runs open-weight models against real banking tasks and picks the right one for each job. Nvidia's NeMo Customizer then tunes that model on the bank's own data. The simple model: a large general-purpose model is a full toolbox carried to every job, when most jobs need one screwdriver. In selected experiments, serving cost fell from €5 per interaction to about €0.03. Why would the smaller model not do worse? Because it is graded on the task, not on general knowledge. The caveat is that those were selected experiments. The one customer-service case Santander described cut costs by more than 50%, not 99%. That is still half. The price of each answer is now the number to manage.
→ Action
Vendor management: Pull per-interaction AI serving costs from model contracts and compare each workload against Santander's €5-to-€0.03 result.
Full story →Read article → from PYMNTS
nCino and Parlay link up to automate SBA lending intake with AI
For the chief lending officer at an SBA lender
The software screens SBA eligibility; your bank still answers for it. Parlay shows $500 million of inquiries, a self-reported 80% cut in file prep, and no funded-loan number yet. Buy it, but test the screening first.
Parlay is software that does the front-desk work of an SBA loan. It takes in the borrower's documents, checks whether the business fits the program, sketches a deal structure, then hands the file to a human banker inside nCino. nCino will resell it as an add-on, and Parlay can still sell direct. The simple model is fewer hours per file: Parlay's customers report up to 80% less manual work per credit-ready file. The flaw is that this figure is self-reported, and the $500 million counts inquiries, not funded loans. The exception that matters is eligibility. The SBA holds the lender responsible for that call, and the OCC and FDIC expect automated credit decisions to be explainable and contestable. So the software can do the paperwork. It cannot sign for it.
→ Action
Model risk and fair lending: Validate Parlay's SBA eligibility screening, audit trails and human override before enabling it in nCino.
Full story →Read article → from The Fintech Times
General AI
Large language models & AI infrastructure
Microsoft calls AI agents insider threats, says frontier models can't be trusted
For the CISO of a regional US bank
Treat every deployed agent like an employee with keys and no judgment: watched by a system its supplier doesn't control, with a stop button. Microsoft agrees, and sells that system, while invested in OpenAI and Anthropic.
Guardian Agents is Microsoft software that watches the AI agents a company runs, inside its Project Perception product. The premise is that an agent is an insider. It holds access to sensitive data and can act, but it was trained to finish tasks, not to follow organizational norms. So Microsoft wraps controls around it. Defender now finds up to 60 agent types that employees bring in, such as Claude Code and Codex. Permissions shrink to the task, and risky moves need human approval. The model is an auditor kept apart from the firm it audits. It is imperfect, because Microsoft funds OpenAI and Anthropic and runs its own lab. The exception is hardware. Nvidia's Open Agent Safety Platform can halt an agent outright. Software can be argued with. A power cord cannot.
→ Action
Information security: Inventory employee-installed agents such as Claude Code and Codex; require human approval before agents act on customer data.
Full story →Read article → from TechCrunch
Hacker uses AI agents to break into John Deere and SharkNinja
For the CISO of a regional bank
The attacker's instructions ran shorter than a text message, under 100 Chinese characters, and still reached more than 100 sites. The weak spot is vendor-run sites nobody watches.
Cairn is an open-source tool. A user sets a start point and an end point, and AI agents work out the steps between, logging progress as they go. One user, according to Reuters, wrote preferred intrusion tactics in fewer than 100 Chinese characters. The agents then ran against more than 100 websites in the US, Canada, Australia and Japan, including John Deere, SharkNinja and Cummins. Each successful break-in took about 23 hours. That is not machine speed. It is a burglar who never tires and pays nothing for each extra house. The limit of this picture: the hits named so far were edges, a vendor-run Cummins site and a standalone SharkNinja support page. For a bank, the gap is less the core than the vendor sites that carry its name.
→ Action
Third-party risk: Inventory vendor-hosted sites carrying your brand and confirm each sends logs to your SOC, with alerts for slow, persistent probing.
Full story →Read article → from The Verge
Washington bans DeepSeek from government agencies
For the chief risk officer at a regional bank
Washington now treats anything built with DeepSeek as something to remove, not just the app. Your bank may run none of it directly and still hold it inside a vendor's model, unlabeled. Find out before an examiner asks.
The US banned Chinese AI firm DeepSeek from most government systems and asked agencies to remove all technology built with it. Deleting an app is throwing out one jar; removing everything built with it means reading every label in the pantry. Why would a bank care if nobody there downloaded DeepSeek? Because models are built from other models, and vendors rarely list the parts. The simple model says a ban covers one product. The imperfect part: 'built with' has no clear edge, and US firms already accuse Chinese rivals of 'distilling' their models, which blurs it further. The exception that matters is your vendors, whose stacks you cannot see without asking. Agencies now must ask. Banks can ask first.
→ Action
Third-party risk: Ask AI vendors in writing whether any model they supply was built on or distilled from DeepSeek.
Full story →Read article → from Semafor
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime