BankingNewsAI Daily Brief ·
New York's AI safety law takes effect, placing OpenAI and Anthropic under state oversight
Banking AI
Financial institutions & fintech technology
UK regulator says banks are trimming AI risk as alarm grows
For the chief risk officer at a US regional bank
The Bank of England's supervisor found no sign UK lenders would amplify an AI shock, and they are trimming anyway, one client at a time. Your bank should know its own AI-borrower list before an examiner asks for it.
The Prudential Regulation Authority supervises British banks. Its head, Katharine Braddick, says a long check of their AI exposure found no sign they would amplify an AI crash. They are cutting anyway. Lenders reread their risk appetite and decide, in her words, "this client, not that client," like a landlord going down the tenant list. The BOE won't use capital rules to force it. So why trim if the regulator sees no danger? Because that finding covers the system, not any one bank's book. PRA and FCA guidance on frontier-AI risk, with governance expectations, arrives this month. Banks moved first. The paperwork follows.
→ Action
Credit risk: Pull your top exposures to AI-sector borrowers and check each one against current risk-appetite limits.
Full story →Read article → from Insurancejournal
BioCatch launches AiRE to detect AI agents in banking sessions
For the head of fraud at a regional bank
Almost half the AI agents BioCatch sees show up at account opening, not at the wire screen. The agent policy you are drafting should start where no customer exists yet: onboarding.
AiRE is a BioCatch feature that tells a bank whether a person or an AI agent is running an online session, and whether that agent seems to work for the customer or for a fraudster. It plugs into existing BioCatch integrations and watches the whole session: synthetic inputs, browser permissions, agent infrastructure, and whether the agent declares itself. It also covers cloud agents like Muse. The easy model treats agents as a payments problem. Yet close to half of detected agent activity came at account opening. Logins, account changes, transfers, wires and faster payments shared the rest. Why would an agent open an account? A new account is an empty drawer you fill later. This is one vendor's customer base, not the market. Detection decides nothing. Your policy does.
→ Action
Fraud operations: Add agent-detection checks and step-up rules to digital account opening before extending agent controls to wires and faster payments.
Full story →Read article → from Finextra
Oracle announces AI-assisted treasury and customer analytics for financial institutions
For the treasurer at a bank running Oracle OFSAA
Oracle now lets a business user type a liquidity question and get a chart back. Your model-validation team has never signed off on a chart nobody can trace to its source.
Oracle has released two analytics apps, Treasury Analytics and Customer Analytics, generally available worldwide. A treasury or finance user types a question in plain English, Oracle Analytics Cloud's AI Assistant draws a chart, and Auto Insights flags patterns nobody asked about. The simple model is one curated data layer that cuts manual reconciliation across ledger, treasury, risk and profitability systems, so answers arrive faster. The flaw is that the assistant decides how to read each question, and that choice sits outside your validated reports. Why would a faster answer be a problem? Liquidity numbers feed funding decisions, and someone will ask how each was built. The exception is that existing on-premises OFSAA keeps running underneath. A calculator on the desk is fine. One that picks its own formula needs a receipt.
→ Action
Model risk: Add AI Assistant liquidity and balance-sheet outputs to the model inventory and require lineage to OFSAA source tables before use.
Full story →Read article → from Oracle
General AI
Large language models & AI infrastructure
Google DeepMind makes Gemini 4 Carbon available to developers and enterprises
For the head of model risk at a regional bank
Carbon costs $2 per million input tokens, and /swarm turns one prompt into a team of agents checked by other agents. Your validation file still assumes one model and one human reviewer.
Gemini 4 Carbon is Google's top model. Developers and enterprises can buy it through the API at $2 per million input tokens and $10 per million output tokens, the same introductory price as Argon. Google says its cost per task is more than 2X lower than the next best external model. The oversight problem is /swarm, a preview in Google Antigravity that turns one prompt into a team. Carbon handles design, coordination and auditing, lighter models work in parallel, and adversarial verifiers test every change. The pitch: machines check machines. The flaw: the auditor and the worker can be the same model. So who validates the auditor? One exception: /swarm is still a preview, while Carbon itself is on sale now. A bank's model inventory has one line per model. This is a whole kitchen staff, and the form has room for one cook.
→ Action
Model risk: Add multi-agent setups like /swarm to the model inventory and require named human validation of the auditing agent.
Full story →Read article → from Deepmind
Anthropic's bug-finding Mythos model is going into Cisco and IBM products
For the CISO of a regional bank
Anthropic kept Mythos behind a vetting queue because it could help attackers exploit flaws. Now it ships inside Cisco licenses at no extra cost. If your red team gets it free, assume attackers' tools are not far behind.
Mythos is an Anthropic model so good at finding software flaws that Anthropic first rationed it to vetted organizations. Cisco's AI Defense, which attacks a customer's AI models, agents and apps to find weak spots, can now call Mythos under existing licenses at no added cost. IBM is adding it to Concert to find flaws traditional scanners miss and to draft fixes. Customers get the results, never the model. It works like a locksmith who tests your doors but keeps his picks. Why widen access to a tool that helps attackers? Because buyers queued and the vendors hold the gate. The gate is not airtight: Anthropic says it built its strongest-ever protections against distillation, meaning cloning. For a bank on Cisco, a stronger test costs nothing. Attackers will not wait for a license.
→ Action
Security engineering: If licensed for Cisco AI Defense, turn on the Mythos red-teaming option and test production AI agents and apps.
Full story →Read article → from Theregister
OpenAI, Anthropic face oversight as New York's AI safety law takes effect
For the third-party-risk officer at a regional bank
From January 1, OpenAI and Anthropic must tell New York about a critical safety incident within 72 hours. Your contract with them may not require them to tell you at all.
The RAISE Act is a New York rule for AI model builders with more than $500 million in annual revenue, such as OpenAI and Anthropic. By January 1, 2027, they must register on a state portal, publish a safety framework, issue model transparency reports, and report critical safety incidents within 72 hours. Fines reach $1 million for a first violation and $3 million after. The simple model: the state now gets a fire alarm. The flaw: it rings at the regulator, not at the bank. Smaller vendors under the revenue line are not covered. California allows 15 days; Illinois adds independent audits. So why should a customer hear second? Put the same report, on the same clock, in the contract. The vendor already has to write it.
→ Action
Third-party risk: Add AI vendor contract terms requiring access to safety frameworks and notice of critical safety incidents within 72 hours.
Full story →Read article → from TechCrunch
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime