BankingNewsAI Daily Brief ·
Three people charged with using AI to impersonate executives and steal $129 million
Banking AI
Financial institutions & fintech technology
Three charged with using AI to steal $129 million by impersonating executives
For the treasury services head at a regional bank
A callback rule assumes the person on the other end is hard to fake. Prosecutors say this ring faked the CEO on video, then supplied the lawyer to call. One victim was a bank.
Prosecutors describe a three-step script. An email reached a company's accounting department, apparently from the CEO, asking for help with a confidential transaction. Next came a phone or video call where AI mimicked the CEO's voice and sometimes face. Then the 'CEO' sent the employee to an attorney at a major law firm, played by a defendant reading a handwritten script titled 'Lawyer Call', who ordered wires to accounts worldwide. The money went into cryptocurrency. Total: about $129 million, with a Rhode Island bank among the victims. Standard advice says verify through a second channel. Here the fraudster supplied every channel, so each check confirmed the last. These are charges, not convictions. The one channel the ring could not supply was a phone number already on file.
→ Action
Treasury services: Require corporate wire callbacks to numbers on file at onboarding, never to contacts supplied in the request email or call.
Full story →Read article → from Cbsnews
UK regulators say enterprises must take blame for their AI agents' mistakes
For the general counsel at a US bank
Four UK regulators, the FCA among them, just said an agent's autonomy does not shrink the deployer's duty. If your agent terms quietly move errors onto the customer, you are betting US regulators will read it differently.
A personal AI agent is software that makes purchases and contacts businesses for a user, without a person approving each step. The CMA, FCA, ICO and Ofcom said jointly that consumer, data, financial services, payments and cyber rules still apply. They added that an agent's autonomy does not reduce the deploying firm's responsibility. Simple model: the agent is a clerk with your office keys. If it opens the wrong drawer, the firm answers. The model is imperfect. The law is unsettled, and these same regulators recently asked for input on how their regimes overlap. The exception: this is UK guidance, not US law. For your bank, the record of what each customer authorised becomes the defence. The agent's cleverness does not.
→ Action
Legal and compliance: Check customer agreements for clauses assigning AI-agent errors to customers, and record the authority each customer grants an agent.
Full story →Read article → from Computerworld
BILL cuts underwriting times from 3 hours to 12 minutes with AI
For the chief credit officer at a small-business lender
BILL's underwriters now finish a case in 12 minutes instead of three hours, and some of the same risk team's agents write model documentation. Your examiners will ask who checked the paperwork that checks the machines.
Robin is a suite of three AI agents that sits in front of BILL's underwriters. Together they send each case to the right underwriter, pull the customer's data into a summary and offer a one-click link to decide. Before, an underwriter pieced each file together from chat messages, email and separate dashboards, like gathering receipts from three drawers. Case time fell from three hours to about 12 minutes. So did the software learn to judge credit? No. It learned to fetch the file, which appears to be where the hours went. A person still clicks. These are BILL's own figures from an earnings call. The exception is the risk team's agents that write model documentation, the records meant to let someone else check the models. Copy the speed, not that.
→ Action
Model risk: Require named human authorship and independent review for any model documentation or validation material drafted by AI tools.
Full story →Read article → from PYMNTS
General AI
Large language models & AI infrastructure
Thinking Machines launches Inkling 2 Small with near-frontier enterprise benchmark scores
For the CTO of a bank weighing on-premises models
A non-Chinese open model now sits a few points behind Claude on agent tests, and Apache 2.0 weights would let you run it in-house. 'Small' means 594 billion parameters, so price the servers first.
Inkling 2 Small is a language model for running agents: software that calls tools, checks its work and reads financial documents. It has 594 billion parameters but uses 17 billion for each word it processes, like a large office where a few clerks handle each letter while you still pay rent on every desk. Why call that small? The name marks its slot in the family, not your server room. On Surge AI's agent test it scored 71.8%, against 77.4% for Anthropic's Claude Fable 5.1. That test runs a fictional PC parts supplier, not a loan desk, so the gap at a bank is unknown. The exception: weights stay closed until safety reviews finish. Once out under Apache 2.0, it becomes the top-rated open model not from China. Then the hardware question starts.
→ Action
Infrastructure engineering: Size GPU memory and server cost to host a 594-billion-parameter model before approving an on-premises Inkling 2 Small pilot.
Full story →Read article → from Venturebeat
Ramp data shows AI spending stalling as Anthropic nears its IPO
For the CFO of a bank renewing AI contracts
Your teams run more AI work each month while OpenAI halves prices and Anthropic cuts its own to keep pace. That is a renewal where the buyer sets terms, from vendors whose forecasts assumed otherwise.
Ramp's data tracks what companies actually pay AI model makers. It shows them sending more work to the models but paying about the same: token spend grew 1.5% across August and September, after 42% in July. In September OpenAI halved API prices for GPT-6 Sol and GPT-6 Luna. Anthropic answered with cuts it says took about $600 million off revenue growth. Buyers also set budgets and route simple yes/no tasks to cheaper models. The simple model: more use, flat bills. It is imperfect, since October spend rose 8.9%. The exception is the top 1% of firms, near $10,000 per employee a month, whose stall drags the total. Why lock in a multiyear price when the price keeps falling? Short terms win. So do lenders who check the forecasts.
→ Action
Procurement: Cap AI vendor contract terms at 12 months and add price-reset clauses tied to published API price cuts.
Full story →Read article → from TechCrunch
Anthropic IPO faces $500 billion valuation hit after tech stocks plunge
For the head of third-party risk at a regional bank
One of three investors put Anthropic near $1 trillion, a third below the expected $1.5 trillion, before the company had even set a price. Your due-diligence file on this model supplier may still treat $1.5 trillion as a fact.
An IPO is a company selling shares to the public for the first time. Anthropic filed on Oct. 26 without a price range, and its Nov. 2 amendment still named no size or valuation. Then Thursday's selloff erased $2.5 trillion from Nasdaq. Three investors told the Wall Street Journal the expected $1.5 trillion is unlikely. One said closer to $1 trillion. The simple model: lower price, weaker supplier. That is imperfect. A valuation is a guess, and the cash Anthropic holds did not move on Thursday. The exception is timing. Reuters sources doubt a November listing, and a delay means outside cash arrives later. So rate the supplier on cash and runway, not headline value. A house is no worse to live in because a neighbor sold cheap.
→ Action
Third-party risk: Reassess Anthropic due-diligence files on cash runway and funding, without assuming IPO proceeds arrive in November.
Full story →Read article → from PYMNTS
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime