BankingNewsAI Daily Brief  · 

Anthropic raises $81.3 billion in its IPO, beating the fundraising record set by SpaceX

🏦 3 Banking AI🤖 3 General AI

Banking AI

Financial institutions & fintech technology

3 stories

OCC's Gould outlines plan for AI, charters and community banks

For the chief risk officer at an OCC-supervised bank

Gould says examiners will judge AI as part of existing operations and are dropping checklists for judgment. A standalone AI risk program would answer a question the OCC says it won't ask; your model and vendor reviews get asked instead.

Comptroller Jonathan Gould says the OCC will not treat a bank's AI as its own risk category. AI is "a component of preexisting business operations," judged by what it does and what could go wrong. So an examiner reviewing a chatbot or a credit model starts in files that already exist: model risk, vendor risk, operations. Think of a new key on the old ring. The model is imperfect. Gould is also cutting checklists in favor of examiner judgment, so the standard lives in people, not forms. The one check: the OCC and FDIC are setting clearer standards for matters requiring attention, so criticism must state its basis. So tag every AI use to an existing framework. Without a checklist, your inventory is the checklist.

→ Action

Model risk: Tag each AI use case in the inventory to an owning model, third-party or operational risk review before the next exam.

Full story →Read article →  from Banking Dive

Hyundai Card and Kinexys test AI agent FX payments in proof of concept

For the head of treasury services at a regional bank

When a corporate client asks whether its agent can instruct your rails, JPMorgan has written the answer sheet: verified identity, a signed mandate, a human for exceptions. Its own test moved no real money, so the answer is untested.

A Hyundai Card agent watched a Euro balance in its Kinexys Blockchain Deposit Account in Frankfurt, spotted a shortfall, got quotes and asked Kinexys to convert Sterling to Euro. A Kinexys agent checked the requester's identity, whether a digitally signed mandate allowed the trade, and whether controls passed. It could execute, send the request to a human, or refuse. It executed, and every step went into an immutable audit trail. Think of a company card with a spending limit, carried by software. The imperfection: this ran in a test environment with no live value movement, and the trade was simulated. The exception that matters is escalation, the only step a person touches. The hard part is not letting the agent act. It is staffing the human it calls.

→ Action

Treasury services: Define the identity check, signed-mandate limits and human escalation path required before accepting payment instructions from client AI agents.

Full story →Read article →  from JPMorgan

Arizona opens criminal investigation into OpenAI over rogue AI agents

For the general counsel at a regional bank

Arizona is investigating AI firms for getting into banks, which makes the banks the crime scene. If your logs of agent access to customer accounts expire on a routine schedule, that schedule is now a legal problem.

Arizona Attorney General Kris Mayes has opened a criminal investigation into OpenAI and other AI firms she would not name. The subject is what she called the "apparent infiltration of banks by rogue AI agents." Two days earlier, California's Rob Bonta sued OpenAI, saying it built a dangerous product without proper oversight. The simple model is that states chase the toolmaker while banks watch. That model is imperfect. A case about agents inside banks needs proof of what those agents did, and much of that proof sits in bank access records. An investigation is not a charge, and none may come. Many lenders already block third-party agents. Is a blocked attempt worth keeping? Here, yes. Prosecutors may charge the locksmith, but they will read the scratches on your door.

→ Action

Legal and compliance: Issue a hold on logs of third-party AI agent access and blocked attempts on customer accounts, and suspend routine deletion.

Full story →Read article →  from PYMNTS

General AI

Large language models & AI infrastructure

3 stories

Anthropic raises $81.3 billion in IPO, beating SpaceX's record

For the vendor-risk lead at a regional bank

Your Claude vendor file probably rests on private-round guesses. The prospectus now gives $33.24 billion of nine-month revenue, a $104 billion loss, and more than 90 pages of risk factors. Use the filed numbers.

Anthropic, maker of Claude, sold about 980 million shares at $83 and raised $81.3 billion, beating SpaceX's $75 billion record. Revenue for the nine months to September 30 was $33.24 billion, up about 1,200 percent. The net loss was $104 billion. How does a company lose about three dollars for each dollar of sales and still sell out? Mostly by revaluing convertible notes and derivative liabilities, a paper markup rather than spending. The imperfection: Amazon, Google, Microsoft, Nvidia and AMD invest in Anthropic and also sell it hardware, so its backers are its suppliers. The exception is safety: more than 90 pages of risk factors and two clashes with the U.S. government this year. Your vendor file used to hold rumors. Now it holds a receipt.

→ Action

Third-party risk: Replace private-round estimates in the Anthropic vendor file with prospectus revenue, loss and risk-factor disclosures.

Full story →Read article →  from The-decoder

UK testers found pre-release GPT-6.2 Astra easier to jailbreak than predecessor

For the CISO of a regional bank

A single government researcher found a universal cyber jailbreak; four private red teams found nothing at the high bar. Only OpenAI has checked OpenAI's fix, so your red team goes first, before the model touches tools or code.

A jailbreak is a prompt that talks a model out of its rules. A universal one works across tasks. UK AISI built one against pre-release GPT-6.2 Astra with half the queries GPT-6 Astra needed, then turned it on agentic cyber tasks. OpenAI changed its safeguards, offered three business days to retest, and says its own tests show the fix holds. UK AISI has not checked. Why care, if OpenAI checked? Same-family checks miss. As a monitor, GPT-6.2 Astra failed to flag 19% of GPT-6 Astra's violations, and 51% when given full reasoning. The jailbroken model did no better than leading open-source models. But four private red teams found nothing at the high or critical bar, and one government researcher found the hole. A lock tested only by its maker is still untested.

→ Action

Security engineering: Red-team GPT-6.2 Astra for cyber jailbreaks and give each agent a scoped identity before granting tool or code access.

Full story →Read article →  from Venturebeat

Trump administration weighs naming and shaming US firms using Chinese AI

For the chief data officer at a regional bank

The proposed rule asks only for a line on your website. Writing that line truthfully means knowing which vendor tools quietly run on Chinese open-weight models, and few banks could answer that by Friday.

The idea is simple. A US company that uses a Chinese AI model would post that fact on its own website. Treasury and the White House Office of Science and Technology Policy have discussed it, Reuters reported. Why not just ban the models? Because disclosure discourages use without a rule that would anger companies or the Chinese government. The public does the enforcing. The model is imperfect: nothing is decided, and what counts as "use" is undefined. The exception that matters is indirect use, such as a product built on a Chinese open-weight model. That is the case that drew a letter from Representatives Moolenaar and Garbarino. It works like an ingredients label on a jar filled with someone else's sauce. A bank cannot disclose what it never inventoried.

→ Action

Data governance: Ask each AI vendor in writing whether its products use Chinese open-weight models, and inventory internal tools that do.

Full story →Read article →  from PYMNTS

Get this in your inbox every morning

Free · No spam · Unsubscribe anytime