BankingNewsAI Daily Brief ·
AI-enabled scam reportedly steals $110 million from French bank Crédit Mutuel Arkéa
Banking AI
Financial institutions & fintech technology
AI-enabled scam reportedly steals $110 million from French bank Crédit Mutuel Arkéa
For the head of payments operations at a regional bank
Arkéa lost about €100m from its own account over instant rails, which carry fewer controls than classic transfers, and no one connected the dots until the third day. Most fraud controls watch customers, not the bank's own account.
Crédit Mutuel Arkéa says attackers took about €100m, roughly one quarter's profit, from its own account. Material from a hacked supplier gave them privileged access. They generated fake IBANs, some at its online bank Fortuneo, and sent payment orders to outside parties over instant payment systems, which carry fewer controls than classic transfers. The money went to legitimate European banks, then to crypto platforms. It ran from Sunday to Tuesday, and crypto specialists saw odd transactions but tied them to no one. The usual model treats instant payments as a customer risk. The chair blames AI, though its role is unclear; stolen supplier access did the work. The exception is the bank's own account, where an inside order looks legitimate by default. The bank was robbed with its own keys.
→ Action
Payments operations: Set instant-payment outbound limits and anomaly alerts on the bank's own operating accounts, including orders initiated through third-party service providers.
Full story →Read article → from Pcgamer
Nubank to expand AI underwriting into US market
For the chief credit officer at a regional bank
Nubank earned $1.1 billion last quarter and says it will treat US credit losses as R&D. Your bank books the same losses as losses. Price your credit book knowing a rival can afford to pay for its lessons.
nuFormer is Nubank's AI model that reads transaction data to decide who gets credit and to catch fraud. It is on its third version. In Brazil it trained on more than a decade of records from a bank about 30% of the country uses as its main account. In the US, Nubank starts with none. Its plan is to begin from bureau files and FICO scores, lend, take losses, and let the model learn from them. Velez calls those losses capital expenditure or R&D. The simple model is a student paying tuition in charge-offs. The flaw: bureau data gives Nubank no edge, since every US lender has it. The exception is money: Nubank earned $1.1 billion last quarter. Why lend to lose? Because the losses buy data. Yours mostly get written off.
→ Action
Model risk: Test any transaction-data credit model for fair-lending disparate impact and explainable adverse-action reasons before using it in underwriting.
Full story →Read article → from Finainews
Starling Bank adds AI-powered voice payments to mobile app
For the head of digital banking at a regional bank
Starling's assistant turns "Pay the plumber £200" into a ready payment, and one tap sends it. Your fraud and dispute rules assume the customer typed the payee. Decide who owns a wrong guess before your customers ask for this.
Starling Bank lets customers pay by telling its app, "Send Mum £20 for lunch." The Starling Assistant prepares the payment, and the customer approves it with a single tap. All retail customers get it by year-end. Small businesses will be able to upload an invoice and tell the assistant to pay it. Think of a clerk who fills out the cheque and hands you the pen. That works while the clerk picks the right Mum. If the assistant picks the wrong plumber and the customer taps without reading, whose mistake is it? A US bank copying this must decide whether one tap on a machine-built payment weighs the same as a typed one, and design the screen for disputes. The tap used to finish the work. Now it is the work.
→ Action
Payments product: Define the payee-confirmation screen and liability rules for assistant-prepared payments before any natural-language payment pilot.
Full story →Read article → from Fstech
General AI
Large language models & AI infrastructure
OpenAI enters cybersecurity market with two Daybreak products
For the CISO of a regional bank
One OpenAI product finds the holes in your web apps; the other writes the pull requests to close them. Your change-management policy assumed a human wrote both. Decide who signs off before the same model works for the other side.
OpenAI sells two security tools under the Daybreak name, both running on its GPT-6 Cyber model. Daybreak Red Team, in Early Access, sends AI agents through a web application's source code and context to map attack paths, check flaws reported by bug bounty programs or scanners, and test patches. Daybreak Defense Factory scans a codebase, writes fixes and tests, and hands developers a ready-to-deploy pull request to approve. The simple model is a locksmith who checks every door nightly and leaves a new lock on the porch. The gap: the lock still needs a person to fit it, and a reviewer facing a pile of tidy fixes tends to approve them. Bank policies assume a human on both ends. Attackers will not need either policy.
→ Action
Change management: Require a named human approver and a separate test run before merging any AI-generated pull request into production code.
Full story →Read article → from The-decoder
Anthropic opens self-hosted Claude Managed Agents in private beta
For the chief technology risk officer at a bank
Self-hosting settles where the agent's secrets and session records live: inside your walls. It also settles who runs the server, the database, the workers and the log review. The vendor-perimeter objection is gone. A staffing question replaces it.
Managed Agents is Anthropic's kit for long, unattended agents: the loop that calls Claude, sandboxes where code runs, a credential vault and a session log. In the private beta your bank runs all of it, including the server, its database and the workers, in its own cloud or data center. Secrets stay outside the sandbox, and a proxy adds them to requests. Each environment gets a host allowlist. Every tool call and result is logged and exported over OpenTelemetry. Simple model: the agent now lives in your building. The flaw: model calls still go out, to Anthropic, Bedrock, Google's Agent Platform, Microsoft Foundry or your own gateway. So does data leave? The prompts do. Zero-data-retention terms decide what Anthropic keeps. Owning the paper trail means owning the reading of it.
→ Action
Technology risk: Add self-hosted agent server, vault, sandbox allowlists and OpenTelemetry session logs to the control inventory and security monitoring review before any pilot.
Full story →Read article → from Aibuzzwire
Thinking Machines releases Inkling 2 Small's 594B weights under Apache 2.0
For the head of model risk at a regional bank
Apache 2.0 means you can run Inkling 2 Small without asking anyone. It also means there is no one to ask. Outside red-teamers had a week; your validation file starts empty.
Inkling 2 Small is an AI model whose full weights anyone can now download, change and run on their own servers. It has 594 billion parameters, but only 17 billion work on each word, so it costs less to run than its size suggests. The simple model: download it, keep customer data in the building, skip the vendor. The imperfect part: the lab calls it wordy, says it sometimes picks the wrong tool, and scores it 30.6 against DeepSeek-V4.1-Flash's 32.6 in its own table. Fine-tuning fixes such flaws. That is the exception: each fine-tune is a new model to validate. The lab's showcase had it train another model alone in 4.5 hours for $426. Cheap to build. The validation file is yours.
→ Action
Model risk: Add self-hosted open-weight models and each fine-tuned version to the model inventory, with validation owners assigned before any download.
Full story →Read article → from Thespecialtynews
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime