BankingNewsAI Daily Brief ·
Anthropic's Claude finds and exploits security flaws in Chrome, Safari, Android and Windows
Banking AI
Financial institutions & fintech technology
Fannie Mae updates its AI and cybersecurity requirements for sellers
For the chief risk officer at a mortgage-selling bank
Fannie Mae renamed AI "super intelligence" and says nothing changes. The real changes are a 36-hour incident clock and vendor compliance landing on your bank by Jan 18. The gap check should start now.
Supplemental Bulletin 26-01, effective Jan 18, resets Fannie Mae's AI and cybersecurity rules. Every AI reference now reads SI, "super intelligence," and the bulletin says the meaning does not change. The real changes sit underneath. A cyber incident starts a 36-hour clock to tell Fannie Mae the event date, the actions taken and whether operations stopped. Any SI use needs risk assessments, data governance, model training and testing controls, and security monitoring. Vendor compliance becomes your job, without the indemnification Freddie Mac requires. Some paperwork disappears: the annual threat risk assessment, fourth-party contingencies, and formal socializing with leadership. Is the bulletin lighter or heavier, then? Lighter on binders, heavier on proof. A 36-hour clock asks whether your vendor answers the phone on a Saturday.
→ Action
Cyber incident response: Add a 36-hour Fannie Mae notice with event date, actions taken and operations status to playbooks and vendor contracts.
Full story →Read article → from Aba
BofA, Citi, Morgan Stanley and Blackstone arrange $63 billion Broadcom AI-compute financing
For the chief credit officer at a regional bank
The first lien in a $63 billion financing of chips leased to Anthropic is backed, dollar for dollar, by Broadcom, the company selling the chips. Copy this structure and you are underwriting Broadcom, not the hardware.
A $63 billion financing lets a new SPV buy Broadcom XPU AI accelerators and networking gear and lease them to Anthropic, with deliveries starting in calendar 2027. Bank of America, Citi and Morgan Stanley arrange the first lien. Blackstone alone commits the committed portion of the second. The simple model: lenders own chips, a tenant pays rent. The flaw: chips lose resale value as newer ones ship, possibly faster than a loan pays down. The exception is Broadcom, which gives residual value support equal to the whole first lien. Why would a seller guarantee what it sold? For the same reason a car dealer quotes a buyback price: it makes the sale happen. The chips are the collateral. Broadcom is the credit.
→ Action
Credit risk: Underwrite AI-compute lease deals on the residual-value guarantor's credit, and stress chip collateral with no residual support assumed.
Full story →Read article → from Morganstanley
Five largest US banks are blocking external AI agents
For the general counsel at a regional bank
Wells Fargo's contract says an authorized agent's instructions are the customer's; one unnamed bank reportedly watched an agent reset a password and move $100,000. If your online terms say nothing about agents, a dispute will write that clause for you.
An AI agent is software a customer lets log in and act for them, including moving money. Five of the largest US banks are blocking outside agents, American Banker reports. So why does Wells Fargo's agreement already deem an authorized agent's instructions to come from the customer? Because the blocks leak. Disguised agents mimic human typing and route around IP blocks. The contract catches what the filter misses, like a lease that bans spare keys yet makes whoever holds one the tenant. The model is imperfect: 'authorized' is the word a dispute will turn on. One unnamed bank, in an unverified account, saw an agent reset a password, change transfer destinations and move $100,000. Standards like Poppy are early. Until they arrive, liability sits where your terms put it.
→ Action
Legal and compliance: Add AI-agent authorization and liability language to online-access agreements, stating who bears losses when customers give agents their credentials.
Full story →Read article → from Pivotnews
General AI
Large language models & AI infrastructure
Anthropic's Claude finds and exploits flaws in Chrome, Safari, Android, Windows
For the CISO of a regional US bank
Vendors get 90 days to patch. The Chrome chain ran code seconds after a page loaded, and one model produced 62 Chrome CVEs. Plan as if attackers will have this too, and cut browser patch windows to days.
Anthropic had an unreleased internal model hunt for flaws in browsers and phones, then write working attacks. In Chrome it chained four bugs: corrupt memory in the JavaScript engine, break out of that engine's sandbox, map the privileged process through a race condition, then run code outside the sandbox via the sign-in feature. A user opens a page. Seconds later the attacker runs code as them. Google has patched 62 CVEs, all critical or high. The simple model says patching closes the gap. That model is imperfect: vendors get 90 days, and your rollout adds more. The exception is Safari, where a chain could read balances and attempt transfers from a signed-in bank session. If one model found 62 flaws in one browser, why would attackers find fewer? Your patches cover your laptops, not your customers' browsers.
→ Action
Endpoint security: Cut forced Chrome and Safari update deadlines on bank devices to days, and confirm every device runs Chrome 156 or later.
Full story →Read article → from Anthropic
Oracle and OpenAI bring ChatGPT agents into Oracle enterprise applications
For the CIO of a bank running Oracle Fusion
ChatGPT plugins will be able to complete approved workflows inside Oracle Fusion, where your ledger and purchase orders live. The same release says it is not a commitment to deliver anything. Settle the model-risk and vendor review before the demo.
Oracle AI Workspace puts ChatGPT plugins in front of Oracle Fusion Applications, NetSuite, Oracle AI Database and OCI. An employee types a goal, and the plugins gather the right Oracle tools and data and complete approved workflows. Picture a new clerk handed keys to the records room and told to act only on signed requests. Oracle says governance, approvals and controls stay in place. The catch is that "approved" means whatever your configuration says, and the clerk's judgment comes from an OpenAI model your bank did not choose. Oracle AI Security Mesh is the exception: there, customers define where testing is authorized and control its scope. So one Oracle contract now brings an OpenAI dependency with it. The ERP was your system of record. It may become a system of action.
→ Action
Third-party risk: Add OpenAI as a fourth party in Oracle Fusion vendor reviews and map which ledger and procurement workflows plugins could execute.
Full story →Read article → from Oracle
Delaware plan would let AI agents run entire businesses
For the BSA/KYC head at a regional bank
The draft bars AI-run companies from the business of banking, not from banking with you. Its one human or corporate member funds it and is shielded from the agent's acts, so your onboarding finds an owner and no decision-maker.
Delaware's draft bill creates an Artificial Intelligence Company, a legal entity whose daily business is run by an AI agent rather than a person. It can sue, hold property and take on debts. One member, a person or corporation, funds it and keeps the profits; the company must log its activities. The member is not liable for the agent's acts unless it underfunds the company, commits fraud or lets it stray from its purpose. The bill bars AICs from the business of banking. Why would a bank care? Because an AIC still needs an account to pay suppliers. Ownership checks ask who owns and who controls. Here the owner is clear and the controller is software, with a log in place of a signer. The ban covers being a bank, not needing one.
→ Action
BSA/KYC: Map how beneficial-ownership and control-person fields would be filled for an entity whose manager is an AI agent.
Full story →Read article → from Arstechnica
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime