BankingNewsAI Daily Brief ·
Citi and Bank of America restrict AI agents' access to customer accounts
Banking AI
Financial institutions & fintech technology
Citi and Bank of America restrict AI agents' access to customer accounts
For the head of digital banking at a regional bank
JPMorgan and Citi keep outside AI agents off the login page. Choose now whether you block, partner or build, before an agent that gets in elsewhere moves your 0.1% deposits to 3.3–5%.
An AI agent is software that logs into a bank account and moves money for its owner, say to chase a better rate. Citi bars third-party agents from logging in or transacting; JPMorgan restricts their access; Bank of America wants a human to approve money leaving. The fear is arithmetic. Checking pays 0.1% on average; Apollo's Torsten Slok says agents could move savings to 3.3–5%. Paying 3% on the $1.6 trillion in non-interest-bearing deposits at JPMorgan, Bank of America and Wells Fargo would cost about $47 billion a year, nearly half their profit. The lock has a gap: read-only open banking access still lets outside services see account information. Blocking the agent slows the money. It does not stop the comparison.
→ Action
Digital banking: Inventory which aggregators and agents hold read-only or credential access to customer accounts, and set a block, partner or build policy.
Full story →Read article → from Incrypted
G7 to war-game AI-driven cyberattacks on global financial architecture
For the head of operational resilience at a regional bank
The drills are on paper; the attacks they model are not. If a scenario pulls your bank in, regulators will want a playbook built for machine speed, not one that waits for a human to convene a bridge call.
G7 finance ministers and central bank governors agreed to run tabletop drills against AI-enabled cyberattacks. Nobody hacks anything. Officials read a scenario and walk through who decides what, and how fast. Each authority writes its own scenarios, from automated liquidity draining to mass data corruption, and may pull in private banks. The simple model: AI runs old attacks faster and wider rather than inventing new ones. The imperfect part is a tabletop moves at the speed of conversation while the threat moves at the speed of code. The exception is shared vendors: one compromised cloud or core processor can freeze several banks at once. Why rehearse slowly for a fast attack? Because the slow part is people. Your playbook will be judged on how many decisions still need a meeting.
→ Action
Operational resilience: Map which liquidity and payment decisions in your cyber playbook need human sign-off, and pre-approve actions for attacks on shared cloud or core vendors.
Full story →Read article → from Dig
FCA publishes good and poor AI practice, calls for AI advice disclaimers
For the chief compliance officer at a US bank
The FCA asks chatbot makers to add warnings, but its practice examples land on the bank whose service sits inside the chatbot. Only 40% of consumers know there is no recourse. Your logo may suggest there is.
The FCA's data chief, Jessica Rusu, published what supervisors saw when they tested firms' AI this year. Good: a chatbot that keeps a live list of vulnerability signals and passes those customers to human advisers. Poor: overstated capabilities, thin testing, governance written after launch. Senior managers are to be personally accountable. The FCA also wants ChatGPT, Gemini and Claude to carry disclaimers on financial advice, because only 40% of consumers know they have no recourse after acting on it. Why ask the chatbot makers rather than banks? Because those companies sit outside its remit. UK rules, not US. But a bank that plugs into those platforms is the regulated name in the conversation, the one with its sign over the door. Accountability flows to whoever can be found.
→ Action
Legal and compliance: Inventory chatbots and LLM-platform integrations; confirm a named senior owner, human handoff for vulnerable customers, and pre-launch testing records.
Full story →Read article → from Fca
General AI
Large language models & AI infrastructure
Anthropic circulates draft federal AI bill mirroring Claude's internal safety policies
For the general counsel at a regional bank
Under this draft, the only federal route to stopping a model your bank runs on is Commerce asking a DC federal judge, over a $1 billion risk. Does your bank examiner count as any other federal agency?
The draft is a federal bill Anthropic has been showing lawmakers, built on its own internal safety policies. Here is the chain. Independent evaluators, registered with Commerce and possibly paid by the labs, review frontier models. If they find an imminent risk of more than $1 billion in damages, or more than 50 deaths or serious injuries, from a bioweapon, a model escaping control, or a cyberattack, they tell Commerce within 72 hours. Commerce then asks a federal judge in DC to suspend the model. No other federal body may. Think of a fire alarm wired to one station. The model is imperfect. A failure below $1 billion, or one that comes some other way, triggers nothing. The exception is preemption of some state disclosure and incident-reporting rules. A model you depend on gets harder to switch off. That cuts both ways.
→ Action
Legal: Map frontier-model vendor contracts for suspension and termination rights, and list which state incident-reporting duties the draft would preempt.
Full story →Read article → from Crowdfundinsider
White House taps FTC to lead AI consumer-protection enforcement
For the chief risk officer at a regional bank
The White House praises AI self-policing, then hands enforcement to an FTC chair already demanding OpenAI's and Anthropic's advertising files. Whatever your bank says its chatbot can do now has to survive that reading.
The White House AI task force has put FTC Chairman Andrew Ferguson in charge of a committee that coordinates AI law enforcement and consumer protection with the Justice Department. The FTC has already sent civil investigative demands to Anthropic and OpenAI for records on security incidents, safeguards, advertising and privacy policies. The simple model: the president signed a "morally binding" light-touch accord, so expect little. That model is imperfect. Light rules plus an active enforcer means policy gets written through cases. So far the demands target AI labs, not lenders. The exception that matters is Ferguson's view that whoever tells a tool to act answers for what it does. A bank that deploys an AI agent is the one doing the telling, like a driver who cannot blame the car. Self-policing holds until someone asks for the files.
→ Action
Marketing compliance: Check every customer-facing AI chatbot claim and disclosure against vendor safeguard records before an FTC demand asks for both.
Full story →Read article → from Aol
OpenAI adds parallel agents to GPT-6 Pro in ChatGPT
For the head of model risk at a regional bank
One analyst, one prompt, several agents each doing a slice of the work. Your model-risk policy assumes a person reviews one model's reasoning. Set workspace model permissions before the gradual rollout reaches your Enterprise seats.
GPT-6 Pro in ChatGPT can now split a task into parts and hand each part to a separate agent working in parallel. The user follows progress in the conversation and can open each completed task to see what that agent did. OpenAI lists GPT-6 Pro on Pro, Business and Enterprise plans, and Enterprise access depends on workspace model permissions. The simple model is one request, one answer, one reviewer. Now one request returns the work of several agents, like a report drafted by a team with one name on the cover. Does inspection fix that? Only if someone opens every task. The exception that matters is the permission setting your workspace admin controls. The rollout is gradual, and that setting is already yours to choose.
→ Action
Model risk: Set Enterprise workspace model permissions for GPT-6 Pro and require review of each agent's completed task before outputs inform decisions.
Full story →Read article → from Neoteo
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime