BankingNewsAI Daily Brief  · 

More banks tell customers who use AI agents that they bear the losses, even if hacked

🏦 3 Banking AI🤖 3 General AI

Banking AI

Financial institutions & fintech technology

3 stories

More banks say AI-agent users are on their own, even if hacked

For the head of consumer digital banking at a regional bank

Five big banks now say a customer who hands an AI agent the keys owns whatever it does, and Citi adds hacks, while exempting its own agents. If your agreement is silent, you keep what they shed.

An AI agent here is software that a customer gives a bank login, so it can read spending, pay bills and move money. JPMorgan and Bank of America told customers last month that whatever the agent does is on them. Citi, PNC and Citizens followed this week. Citi goes further. If the tool leaks the password or gets hacked, the thief's transactions are the customer's too. The simple model is a spare house key handed to a stranger, where the homeowner answers for the burglary. It is imperfect, because the customer cannot see how the agent stores the key. The exception is Citi's own agents, which carry no added liability. So who gains when outside agents become risky to use? The bank that offers its own.

→ Action

Legal and compliance: Compare your online user agreement's AI-agent and compromised-credential clauses against Citi's update before customers or examiners ask.

Full story →Read article →  from TechCrunch

Visa expands Agentic Ready to acquirers, payment providers and merchants

For the cards and acquiring chief at a regional bank

Fiserv, Stripe and 140-plus participants now test how an AI agent's card payment gets recognized, beside 175+ issuers. More join through 2027. A bank that waits will process those payments under rules it never tested.

Agentic Ready is Visa's testing program for card payments that an AI agent starts for a shopper. More than 175 issuers in 45 markets joined first. Now more than 140 acquirers, merchants and payment providers join too, including Fiserv, Global Payments, Stripe, Adyen and PayPal. They run real agentic network token transactions through authorization and processing to work out how a merchant recognizes and authenticates a trusted agent. Simple model: the token is a key cut for one agent and checked at both ends. The flaw: agent identity standards are still being explored, so the key exists before anyone agrees on the lock. The exception: it all runs on existing payment infrastructure, so these transactions reach your bank whether you joined or not. Why test early if the rules come later? Because the testers write them.

→ Action

Merchant acquiring: Ask Fiserv or Global Payments whether your merchant portfolio can join Agentic Ready testing of agentic network tokens in 2027.

Full story →Read article →  from Visa

UK regulators publish effective practices for AI cyber risk and resilience

For the CISO of a US regional bank

UK regulators say AI lets attackers exploit flaws in hours or days, not weeks or months, yet they still want humans on key defence decisions. It binds nobody. It is still a free answer key for your next exam.

The Bank of England, PRA and FCA have published effective practices for cyber risk from AI. They are not formal expectations; they add detail to cyber guidance. AI reads source code and maps attack paths cheaply, so flaws once exploited in weeks or months now fall in hours or days. Discovery stops being the bottleneck. Repair becomes it: find affected assets, test the fix, get a change window, coordinate third parties, deploy, verify. Why not let AI run that chain? It can give different answers on repeated runs, so humans keep key decisions. Firms scale this to their reliance on AI and to how much access those systems have. Locks now get picked in an afternoon; the locksmith still books a month out. The attacker got the speed. The signature still needs a person.

→ Action

Information security: Measure days from vulnerability disclosure to verified patch on critical systems, and name who approves emergency changes.

Full story →Read article →  from Bankofengland

General AI

Large language models & AI infrastructure

3 stories

Amazon says stolen AI API keys power agents hunting for more

For the CISO of a regional US bank

Your developers' AI gateway is a drawer of model-provider keys, and attackers now spend those keys to run agents that find the next drawer. The FBI saw some intrusions reach privileged access in under three hours.

ARTEX is an open-source penetration-testing tool. It turns plain-English instructions into commands run by AI agents, using outside models. Amazon's decoy sensors caught records of agents scanning exposed LiteLLM, Dify and Langflow installs, trying default master keys and the Langflow flaw CVE-2025-3248, then pooling the provider keys that worked. Those keys pay for the model calls behind the next scan. The victim funds the search. A human attacker sends dozens of commands an hour; CrowdStrike saw an agent send about 1,500. The caveat: Amazon did not watch the attacks and could not verify outcomes. The exception that matters is banking. ARTEX attacks on South Korean financial firms took personal data on around 67,000 people. Who at your bank owns the gateway a team set up for a pilot?

→ Action

Security engineering: Inventory internet-facing LiteLLM, Dify and Langflow installs, rotate default master keys, and patch Langflow CVE-2025-3248.

Full story →Read article →  from Venturebeat

Google DeepMind introduces Gemini 4 Neon at $3/$15 per million tokens

For the chief technology officer at a regional bank

A model ranked #3 on an independent index now reads a 1M-token file for $3 per million tokens and re-reads it for $0.15. Your incumbent vendor's quote for KYC and credit-memo work now has a number to beat.

Gemini 4 Neon is Google's new model, sold by the token: $3 per million in, $15 per million out. One request can hold 1 million tokens of spreadsheets, PDFs and charts. The line that matters is cached input. Send the same file again and the charge falls to $0.15 per million, a 95% discount. It is like paying full fare once to enter a records room, then pocket change for each return. A KYC file questioned many times costs little more than one read. Does the ranking hold up? Independent indexes put it #3 and #2, but the low fabrication claim comes from Google's own testing. Output costs five times input, so price your pilot on memo drafting, not file reading. Reading is now cheap. Writing is where the bill sits.

→ Action

Vendor management: Add cached-input and output-token prices to the current model RFP, and require bidders to quote cost per KYC file review.

Full story →Read article →  from Deepmind

Microsoft unveils new Copilot, Autopilot agent and agent controls at Ignite 2026

For the chief operating officer of a microsoft-standard bank

Microsoft now ships an agent that moves work forward without asking at each step, with its own Entra identity. Your access, records and model-risk rules assume a person clicks send. Fix that before staff switch it on.

Microsoft is folding Office into Copilot and adding Autopilot, a long-running agent that handles recurring tasks and moves projects forward without a prompt at each step. Copilot Code lets anyone build dashboards, trackers and apps in natural language. The simple model: software that once waited for a click now acts, under its own Entra identity, with Agent 365 policies enforced on each device. The flaw in that model is that an identity records who acted, not whether the action was right. Why worry if Microsoft ships the guardrails? Because guardrails enforce the rules you write, and most bank policies were written for people. The exception is BNY, which worked with Microsoft to complete the first agentic negotiation of a U.S. Treasury trade. A house key opens the door for whoever holds it. Name the holder.

→ Action

Model risk and records management: Add Autopilot agents and Copilot Code apps to the model inventory and retention schedules, each tied to an Entra agent identity.

Full story →Read article →  from Microsoft

Get this in your inbox every morning

Free · No spam · Unsubscribe anytime