BankingNewsAI Daily Brief ·
US bank regulator issues its first-ever AI guidance for examiners
Banking AI
Financial institutions & fintech technology
US bank regulator issues first-ever AI guidance for examiners
For the chief risk officer at a national bank
Under the OCC's new exam guidance, a plain agent that can move money may draw more scrutiny than your most sophisticated internal model. Write down what each such agent may do without a human before examiners ask.
The OCC has told its examiners, for the first time, how to look at banks' AI. Judge a system by what happens when it is wrong, not by how clever it is. Banks need no separate AI risk parameters. Examiners weigh how the tool is used and the bank's size and complexity. So why would a simple agent count as riskier than a sophisticated model? Because the agent can initiate transactions and the model stays internal. A clerk holding the vault key matters more than a genius in the back office. The guidance says could, not will, and it is not legally binding, though banks typically follow such guidance closely. Higher-risk uses may need tighter autonomy limits or direct human oversight. Sophistication was never the exam question. Reach is.
→ Action
Model risk: Inventory every AI agent that can initiate transactions and document its autonomy limits and human-oversight checkpoint.
Full story →Read article → from Y94
Morgan Stanley lets advisors build their own AI agents
For the COO of a bank's wealth management unit
Morgan Stanley reviews advisor-built agents after they are in use, not before, and advisors have attempted nearly 50,000. Your choice is whether model risk can live with review-after-use, or whether you block the builders.
Morgan Stanley lets advisors write a plain-English prompt and get a working AI agent, such as one that pulls every outgoing flow above a set amount across their book each day. Advisors have made nearly 50,000 attempts. Most die in early testing, and some are used twice and abandoned. The control model has three tiers. Orphans get no review. Agents a team adopts get checked for accuracy and completeness after use. Firmwide candidates get more testing. That is triage, not prevention. Why not review everything first? Because 50,000 pre-checks would be impossible. The exception is data: an agent used twice still read client accounts twice. The gate that matters is what an agent can touch, not whether it survives. Review sorts output; access sorts risk.
→ Action
Model risk: Set data-access entitlements for user-built AI agents before build, and inventory agents in repeated use for post-use validation.
Full story →Read article → from Financial-planning
HSBC opens transaction banking data to corporate AI tools
For the treasury-services head at a US bank
HSBC now lets a client's own AI read account and transaction data through MCP, inside permissions the client already set. Your corporate clients will ask why their agent can read HSBC's accounts but not yours.
HSBCnio is HSBC's new corporate banking platform, and its key feature is a door for machines. Through the Model Context Protocol, a client's own AI tool can read authorised account and transaction data, within existing access permissions, and do nothing else. It is like handing the accountant a key to the records room but not the checkbook. The treasury team asks its agent for a liquidity forecast, a variance analysis or a cross-border cash view, and the agent pulls the numbers itself. Why stop at read-only? Because an agent that only reads can be wrong cheaply. The exception: HSBC plans to bring more transaction services into MCP over time. Once agents can act, your entitlement model is the only thing between a prompt and a payment.
→ Action
Treasury-services technology: Map each corporate user entitlement to API scopes so a client's AI agent inherits read-only, per-account limits by default.
Full story →Read article → from FinTech Global
General AI
Large language models & AI infrastructure
Best frontier AI scores 12% on GDPx, zero on finance and legal
For the cto of a bank weighing agent vendors
Vendor decks quote benchmark scores; on GDPx, every model scored zero on finance and legal work done the way analysts do it. Approve agent autonomy in those functions only after a vendor passes your own task, not theirs.
GDPx is a SemiAnalysis test that puts AI agents inside a simulated job, not a single prompt. The finance task gives a model about 4,000 files from a reconstructed $880 million private-equity deal and up to eight hours, the work an analyst does in a deal room. Across 206 tasks and 13 models, the best, Claude Opus 5.5, scored 12%. On finance and legal, every model scored zero under pass-or-fail grading; the test switched to partial credit to rank them. Why do vendor decks look better? Older tests like GDPval fit the job into one prompt, which is like grading a cook on the recipe card. One caveat: safety filters hold back Opus 5.5 and GPT-6 Astra in three domains, none of them finance or legal. The zero stands.
→ Action
Technology risk: Require vendors to run agent pilots on your own finance and legal files, graded pass or fail, before granting autonomy.
Full story →Read article → from Techtimes
Anthropic builds open-source system to prove which AI model produced an output
For a third-party risk manager at a regional bank
Your AI contracts name a model and take the vendor's word for it. FlashProve could check that word, but no lab plans to deploy it, so write the verification right in now.
FlashProve is a proof of origin for AI answers, open-sourced by Anthropic. A lab publishes a fingerprint of each model, and any output can be checked against it on a local device, the way an email signature confirms the sender. The fingerprint does not reveal the weights, so closed-model makers could publish it without giving away the product. Who needs that? Anyone paying for a model by name. The researchers name the cheat: a provider quietly swapping in a cheaper model to cut costs, or a pricier one to inflate performance results. Research is ongoing, and no lab, Anthropic included, plans to deploy it. So the control exists as code on GitHub, not as a service. A contract can ask for fingerprints before anyone ships them. It should.
→ Action
Third-party risk: Add AI vendor contract terms naming the model served, requiring notice before any swap and published model fingerprints once available.
Full story →Read article → from TechCrunch
Oracle drops $7 billion Tencent AI chip lease deal under US pressure
For the sanctions officer on a corporate-banking credit desk
The data centres were in Southeast Asia; the customer was in China, and Washington judged by the customer. A loan repaid from lease income paid by Chinese tenants can lose that income before the lease is even signed.
Oracle planned to lease Tencent up to 100,000 advanced AI chips over five years, a deal worth about $7 billion. The chips would sit in Southeast Asian data centres, because US export rules bar shipping them straight to China. Trump administration pressure ended the deal before signing, the FT reported. The simple model says export rules follow the hardware, so hardware kept offshore is safe. That model no longer holds. The test is who uses the hardware. Beijing also pushed Tencent toward domestic chips, so pressure came from both sides. Nvidia has reportedly told cloud providers that at least 51% of their services must go to non-Chinese customers. For a lender, a data centre's value is its tenant list. Check the passports on it.
→ Action
Credit risk and sanctions: Flag data-centre loans whose lease income depends on Chinese tenants; test them against a 51% non-Chinese customer threshold.
Full story →Read article → from Chipdispatch
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime