BankingNewsAI Daily Brief ·
OpenAI launches dots, always-on AI agents that work around the clock
Banking AI
Financial institutions & fintech technology
FSB finalises 12 sound practices for responsible AI adoption
For the chief risk officer at a regional bank
The FSB calls these practices a toolkit to tailor, not a rule. Examiners tend to read a list a board is told to reference as the list the board will be asked about. Map to all 12 now.
The Financial Stability Board, which coordinates authorities from 24 jurisdictions, has finalised 12 practices for using AI in financial firms. Four cover governance: board oversight, defined roles, AI risk in the risk framework with documented use, and continuous learning. Eight cover the lifecycle, from materiality assessment through explainability and human oversight to third-party AI dependencies. The simple model is a voluntary checklist. So why would a voluntary list matter? The FSB strongly encourages boards to reference it, and the national authorities that make up the FSB are the ones banks answer to. The honest caveat: firms may tailor it to their risk profile. The exception is practice (v). Without an inventory of AI use cases rated for materiality, there is nothing to tailor. An inspector asks for the floor plan before checking the wiring.
→ Action
Model risk: Build an inventory of AI use cases, rate each for materiality, and map its documentation to the FSB's 12 practices.
Full story →Read article → from Fsb
FIS makes its AML investigations agent available to any US bank
For the BSA/AML officer at a US bank
Amalgamated's numbers are projections from a bank that helped build the agent, not results. If case review time falls 98 percent, your examiner will ask what the analyst still checks before signing the SAR.
FIS now sells an AI agent, built with Anthropic on Claude models, that gathers evidence for AML alerts, recommends a disposition and drafts the SAR narrative. Any US bank can buy it, FIS client or not. Nothing closes without analyst confirmation, and each statement links to its source record. Amalgamated Bank, a design partner, projects cuts of more than 90 percent in Level 1 review time, about 98 percent in Level 2 investigation time and more than 80 percent in SAR drafting. These are projections, not results. Native integration covers only Horizon and Systematics. So who wrote the SAR? At 98 percent, the analyst keeps two minutes of every hundred, which is time to sign, not to read. Your validation must show the signature still means review.
→ Action
Model risk: Add the FIS AML agent to validation scope, testing sampled SAR narratives against their linked source records before any pilot.
Full story →Read article → from Fisglobal
PNC expands in-house AI capacity while keeping hyperscaler models
For the CFO of a large regional bank
PNC is turning a cloud compute bill into depreciation on its own GPUs, yet still rents frontier models from hyperscalers. Your budget faces the same split: which half of the AI stack you own, and which you keep paying for.
PNC is putting GPU chips into its own rebuilt data centers so it can run AI workloads, including open-weight models, on machines it owns, while still reaching frontier models through hyperscalers. Rented compute is a bill paid to a third party. Owned compute is paid up front and then lands as depreciation, which PNC says is already rising. The simple model is to own the steady work and rent the frontier. It is imperfect, because the strongest models still sit with outside providers, so vendor dependence shrinks rather than ends. The exception is the payback. The $1.4 billion across 171 opportunities is addressable spend, not savings, and the 40 points of operating leverage by 2030 is a forecast. Owning the garage only pays if you drive the car.
→ Action
Finance planning: Model three-year cost of owned GPUs, including depreciation, against current third-party AI compute spend before the next budget cycle.
Full story →Read article → from Tokenpost
General AI
Large language models & AI infrastructure
OpenAI introduces dots, always-on AI agents that work 24/7
For the CISO of a regional bank
OpenAI's dots run 24/7 on their own cloud computer and plug into over 4,000 apps through plugins. Your data-loss controls watch employees' laptops. The agent's computer is not one of them.
A dot is an OpenAI agent, powered by GPT-6 Astra, that keeps working after the employee goes home. It runs on its own cloud computer, reaches over 4,000 apps through plugins, and takes instructions in ChatGPT, Slack or Teams. Think of it as an assistant holding a copy of an employee's keys. The model is imperfect. The keys sit on a computer the bank does not own, and they are used around the clock with no one at the desk. Why does an assistant need its own computer? Because doing work, not just answering, needs somewhere to run. The exception is memory. A dot learns how its user thinks and what good looks like. Access reviews built for people will not catch a helper that grows into the job.
→ Action
Identity and access management: Block dots and OpenAI plugin connections in ChatGPT, Slack and Teams admin consoles until agent-access and data-loss policies exist.
Full story →Read article → from OpenAI
Proofpoint agrees to buy Varonis for $6.3bn in cash
For the CISO of a bank running both tools
Your communications-security vendor and your data-permissions vendor will soon be one Thoma Bravo-backed company that paid 53 times free cash flow and plans to merge products fast. Fix roadmap and pricing terms before that plan reaches your renewal.
Varonis sells software that finds sensitive files, tracks who can open them and, more recently, what AI systems can reach. Proofpoint, owned by Thoma Bravo, agreed to buy it for about $6.3bn in cash, or $49.75 a share, 52% above the price before sale talk began. That is roughly 53 years of Varonis' current free cash flow. A buyer paying that does not leave two products alone. Proofpoint says it will merge the technology, align sales and push one platform quickly to more than 16,000 enterprises. The simple model: one vendor, one console, fewer contracts. The imperfection: merged products often drop features customers relied on. The exception is a bank already running both, which has bargaining power now and less after close. Someone pays for the premium. Usually it is the customer, at renewal.
→ Action
Vendor management: Add roadmap, price-cap and product-retirement clauses to Proofpoint and Varonis renewals before the merger closes.
Full story →Read article → from FinTech Global
Anthropic gives METR researchers employee-level access to assess its models
For the head of third-party risk at a bank
One AI vendor now lets six outside researchers sit at its desks and publish without its veto, the deal regulators impose on the largest banks. Ask every other model vendor in your inventory whether it accepts anything close.
Anthropic has given six researchers from METR, a nonprofit that tests AI models, company laptops, office space and internal tools roughly on par with its own risk teams. METR publishes its own findings on Anthropic's risks, incidents and practices, and Anthropic cannot veto them. The simple model is a resident examiner, Anthropic's own comparison. Where it is imperfect: Anthropic can force redactions of customer data, privileged material and model architecture. Why accept a redaction right? Because METR may say publicly whether each redaction was justified, so a black bar becomes evidence too. The exception: METR alone sits in the top tier, and bottom-tier reviewers like SecureBio need Anthropic's approval to publish, so read those as vendor-cleared. OpenAI pledged the same. A pledge is not a laptop.
→ Action
Third-party risk: Add AI vendor due-diligence questions on whether outside evaluators get internal access and can publish without vendor approval.
Full story →Read article → from TechCrunch
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime