BankingNewsAI Daily Brief ·
Apollo's record $114 billion in AI lending begins to rival the biggest banks
Banking AI
Financial institutions & fintech technology
State Street urges responsible AI as banks race to deploy agents
For the chief data officer at a regional bank
State Street went from about 150 to over 350 AI use cases and is weighing its own agent gateway because no shared protocol exists. Help write the standard, or build a door you will rebuild later.
State Street wants its AI agents, software that does tasks, to swap data with other firms' agents without letting any outside agent inside. Why not let them talk directly? That, says data chief Manoj Bohra, puts a foreign object in your company. So it is considering a gateway, the only route in or out, where each agent has a predefined identity and only the entitlements one customer and process need. Think of a lobby desk issuing single-floor badges. The flaw: if every bank builds its own desk, counterparties need a badge for each. Bohra says a shared standard would be ideal, and the industry is far from one. Use cases went from about 150 to over 350 in a year. The doors are going up before the rules.
→ Action
Data governance: Inventory which vendor or counterparty AI agents touch bank data, and require each a defined identity and least-privilege entitlements.
Full story →Read article → from Scmp
Sentra raises $150 million in round led by Capital One Ventures
For the CISO of a regional bank
Capital One now sits on the board of a company whose pitch is that each AI agent is a new identity. If your access reviews list only people and service accounts, your agents are not on the list.
Sentra sells software that finds sensitive data across cloud, SaaS and on-prem files, then maps who and what can reach it: people, service accounts and AI agents. Capital One Ventures led a $150 million investment and takes a board seat. Why not just watch each agent request as it happens? Agents read at machine speed, so Sentra secures the data beforehand and revokes access when behavior looks odd. The simple model is that an agent is an employee who never sleeps. The imperfect part is that many agents inherit access to data that was never fully classified or governed, so nobody granted it on purpose. That inherited access is the exception that matters. A bank whose access reviews list only people may not know how many keys its agents already hold.
→ Action
Identity and access management: Add AI agents and their inherited permissions to access reviews and least-privilege checks before any agent reaches production.
Full story →Read article → from PR Newswire
Embat integrates J.P. Morgan API to slash bank onboarding times
For the treasury services head at a regional bank
A corporate client can now link its J.P. Morgan accounts to Embat in seconds, while linking yours may still take months of certificates and approvals. Guess which bank's balances the CFO checks first.
Treasury software like Embat pulls a company's bank balances, transactions and payments into one screen. To do that, it needs a trusted connection to each bank. Until now, the company got certificates from a third-party authority, then sat through implementation work, data checks and approvals that could delay onboarding by months. J.P. Morgan Payments' Banking Authorisation Service API cuts that out. Embat uses its own certificates, the client signs legal papers, and accounts link within seconds. One security setup covers many subsidiaries. Factorial and Cazoo are first. The caveat: Embat itself says 'days,' not seconds, to go live, and legal sign-off still takes time. The key used to come from a locksmith across town. Now one bank hands it over at the counter, and every other bank looks slow.
→ Action
Treasury services: Time corporate TMS and ERP connection onboarding end to end, and list each certificate and approval step that could be removed.
Full story →Read article → from FinTech Global
General AI
Large language models & AI infrastructure
OpenAI pipeline produces 948 Linux root exploits for $1.2 million
For the CISO of a regional bank running Linux
A working path to root on Linux now costs about $1,279 in API credits, against a historical $20,000 bounty. Any bank server with a low-privilege login and a lagging kernel should be treated as already rooted.
OpenAI built a pipeline that scans the Linux kernel, reproduces bugs, writes working exploits and drafts patches, with a human reviewing the output. One October run on a stock desktop distribution produced 948 exploits that lift an ordinary user to root, at $1,279 each. Pwn2Own historically paid $20,000 for one. Why publish that? Because the same run yields patches, and others patched or staged patches for over 100 such flaws within 10 days. The caveat: bounty prices are not black-market prices, and the model behind this is not yet released. The exception is the server behind on patches, where OpenAI says a low-privilege login equals root. A good lock helps little once copies of the key are cheap. For a bank, the kernel patch cycle is now the lock.
→ Action
Infrastructure security: Shorten Linux kernel patch SLAs and isolate low-privilege workloads with gVisor or Firecracker sandboxes on shared servers.
Full story →Read article → from OpenAI
Apollo's record $114 billion AI lending begins to rival banking giants
For the chief risk officer of a corporate bank
Apollo now lends at Citi's scale against chips whose value Ares's CEO says he cannot underwrite. Banks already fear single-name Big Tech exposure; competing for these deals adds more, secured by an asset that ages like a laptop.
Apollo lent a record $114 billion last quarter, 50% more than the quarter before, and over $350 billion in 12 months, in line with HSBC and Citigroup. Much went to Nvidia, Anthropic and SpaceX/xAI, secured against chips. The simple model: chips are valuable, so they make good collateral. The flaw: a chip ages like a laptop, not a building, and Ares's CEO says he cannot underwrite how fast its value falls. So why do banks leave these deals to a fund? Many already fear relying too much on single Big Tech names. Apollo's own economist says AI has paid off zero outside tech. The fund is not taking risk banks missed. It is taking the risk banks already hold.
→ Action
Credit risk: Aggregate exposure to Nvidia, Anthropic and xAI across loans, commitments and counterparties, and set haircuts for chip collateral.
Full story →Read article → from PYMNTS
AMD's AI data center guarantees reach $8.5 billion as it challenges Nvidia
For a credit officer lending to data-center developers
AMD now backs $8.5 billion of customer leases and debt, plus $13.3 billion of leases not yet started, against $15.8 billion in cash. If your borrower's tenant credit rests on AMD, so does everyone else's.
AMD is co-signing its customers' data-center rent. It guarantees about $8.5 billion of leases and debt. It has also signed $13.3 billion more in leases that haven't started, which it plans to pass to partners. Why would a chipmaker pay a customer's rent? Because the building is where its chips go. In the Core Scientific deal, AMD leased 377 of 529 megawatts itself and backed a cloud provider on the other 152, taking warrants on up to 30 million shares. The simple model treats AMD as a strong guarantor behind a weak tenant. The flaw: both can fail for the same reason, if AI demand slows. No partner has defaulted yet. So a lender should count AMD support as correlated, not independent. A co-signer who also sells the furniture is not a neutral co-signer.
→ Action
Commercial credit: List every data-center loan with AMD or Nvidia lease or debt guarantees and total the exposure by guarantor.
Full story →Read article → from Yahoo Finance
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime