BankingNewsAI Daily Brief ·
Citi says customers are responsible when AI banking agents go rogue
Banking AI
Financial institutions & fintech technology
Citi says customers are responsible when AI banking agents go rogue
For the general counsel at a regional US bank
Citi now treats a hacked agent's transfers as customer-authorized; JPMorgan and Bank of America have similar policies. Federal rules say contracts cannot waive fraud protection. Your terms must pick a side before your first agent dispute does.
An AI agent is software that a customer hands a banking login so it can act for them: read records, move money. Citi's new terms say anything done with those credentials counts as customer-authorized, even if the agent ignores instructions, exceeds its permissions or leaks the login to someone else. Citi also need not check whether the agent followed the prompts. JPMorgan and Bank of America adopted similar policies in October. The simple model: lend a contractor your house key, and the break-in is on you. The imperfection: federal rules require banks to investigate unauthorized transfers, and the CFPB says agreements cannot waive that. The exception is PNC, which bans sharing passwords and one-time codes and allows only bank-approved connections. A contract can assign blame. Only a control can prevent the loss.
→ Action
Legal and compliance: Map digital-banking terms on third-party agent credentials against unauthorized-transfer investigation duties, and decide whether to ban password and one-time-code sharing.
Full story →Read article → from Yahoo Finance
FCA launches AI Spotlight Observatory to track financial services AI adoption
For the UK chief risk officer at a US bank
The FCA has published a free map of how 200-plus firms use AI. It is not a data request, but UK supervisors now hold a peer benchmark, and your AI inventory will be read against it.
The AI Spotlight Observatory is a free website run by the UK Financial Conduct Authority. It collects real examples of AI use from more than 200 firms across 150 functions, from financial crime tools to agentic payments. CFTE built the taxonomy; NayaOne built the search and filters. Why would a regulator build a catalogue instead of a rulebook? A catalogue sets a norm without a consultation. The simple model is a showroom where firms learn from peers. It is imperfect: these are examples firms chose to show, not a census. The exception is the supervisor, who can now open a page of peer uses and ask why yours differ. A neighbour's tidy garage makes yours look worse. Your AI inventory was always a record. Now it has a comparison.
→ Action
Model risk: Map your UK AI use-case inventory to the Observatory's taxonomy and attach governance evidence to each use case.
Full story →Read article → from Fintechfutures
BMO demonstrates quantum computing milestone for AI and machine learning
For the chief data officer at a regional bank
BMO's quantum feature selection on IBM hardware matched classical methods using one to two orders of magnitude fewer data samples. Matching is not beating, so the budget case rests on data you lack, not accuracy you need.
Feature selection is the step before a model runs: from a long list of variables, pick the ones that matter. BMO, working with the Democritus University of Thrace on IBM's Nighthawk r2 processor, treated each variable as a player in a quantum game and let the machine rank them. The ranking matched traditional techniques while using one to two orders of magnitude fewer data samples. That is like pricing a store's stock from a tenth or a hundredth of the receipts. The catch: the result is a match, not a win, and BMO says it is in the early stages. So why would fewer samples matter to a bank buried in data? Because confirmed fraud cases are the samples nobody has enough of. That is where a quantum pilot earns its budget, or doesn't.
→ Action
Data science: List fraud and risk models limited by scarce labeled samples; make them the required test case for any quantum pilot proposal.
Full story →Read article → from Fstech
General AI
Large language models & AI infrastructure
Anthropic IPO heightens Washington's scrutiny of the tech sector
For the corporate secretary at a listed bank
Four senators want every public company to disclose more about AI, not just the ones selling it. Your bank's filings probably mention AI in scattered places. Better to find them before Congress decides what the words must mean.
The bill is simple to describe. A bipartisan group of four senators wants publicly traded companies to disclose more about AI. Anthropic's IPO is the trigger: a big AI company now answers to public shareholders, and lawmakers' worries about model abuse and market volatility have a listed company to point at. The simple model says this is about AI makers. It is imperfect, because the bill as described covers publicly traded companies, and a listed bank is one. The gap that matters is detail. The article gives none on what must be disclosed, or when. Elizabeth Warren, the top Democrat on Senate Banking, wants quick action, and Democrats are preparing to take control of Congress. A bank using AI vendors is like a tenant asked to list every key cut to its office. Easier to count them before the landlord asks.
→ Action
Legal and compliance: Inventory every AI mention in the latest 10-K, including vendors and risk factors, and assign one owner for disclosure language.
Full story →Read article → from Semafor
Amazon says attackers' AI agents exploited a zero-day in 11 minutes
For the ciso of a regional bank
Attacker agents needed about 11 minutes to find and use the flaw. The merged fix has waited for a release since Oct. 19. For open-source tools you expose, the patch is the slow part, so cut exposure first.
MONAI Label is an open-source medical imaging app. Attackers could trick its file-upload feature into writing files outside its storage folder. In a later session they turned that into running code and full control of the server. Amazon read their prompts on an unsecured Cairn server, which coordinates AI agents. The agents had one goal, "obtain shell access," and reached a zero-day in about 11 minutes. The fix was merged Oct. 19. The newest release is still 0.8.5, from Nov. 25, 2024. Why not patch? There is nothing to install. The lock is fixed at the factory, and your door has the old one. The campaign hit universities, not banks. Authentication is off by default, and ANNOTATOR users can exploit the flaw even with it on. The release is now the slowest step. Block network access first, then check for earlier break-ins.
→ Action
Vulnerability management: Inventory internet-facing open-source servers with authentication off by default, and restrict network access now instead of waiting for releases.
Full story →Read article → from Venturebeat
OpenAI's GPT-6.2 Sol and GPT-6.2 Astra now available in Microsoft Copilot
For the cio of a bank running microsoft 365
Two new OpenAI models are rolling out in Copilot, Sol inside Word, Excel and PowerPoint, reading staff files and chats. Your model-risk team has likely validated neither. The admin center switch decides which comes first.
Microsoft is adding two OpenAI models, GPT-6.2 Sol and GPT-6.2 Astra, to Copilot. Sol is rolling out worldwide in Copilot Chat, Cowork, Studio, Excel, PowerPoint and Word. Astra goes to Cowork and Studio. Work IQ feeds both models the user's files, meetings, chats and business data, limited to what that person can already open. So where is the risk, if nothing new is exposed? Permissions control what the model reads, not what it writes. It is like handing a new temp the same office keys as the employee. No extra doors open, but someone unvetted is now drafting the memo. The exception: Frontier users also get Astra inside the Office apps. Admins control access in the Microsoft 365 admin center. The vendor sets the rollout schedule. Validation is still yours.
→ Action
IT and model risk: In the Microsoft 365 admin center, restrict GPT-6.2 Sol and Astra access for production users until model validation is complete.
Full story →Read article → from Microsoft
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime