BankingNewsAI Daily Brief ·
House votes to modernize bank oversight and end penny production
Banking AI
Financial institutions & fintech technology
Alchemy unlocks AI agent purchases via AgentCard
For the card payments head at a regional bank
Your cardholders may be able to let AI agents spend against existing Mastercard accounts without a new card account. Test whether your fraud, dispute and consent rules can distinguish an approved agent purchase from an unauthorized one before agent use scales.
Alchemy said AgentCard will support Mastercard payment credentials through an integration with Mastercard Agent Pay. Developers can give an AI agent an email address, phone number, stablecoin wallet and one-time-use Mastercard payment credentials through a single CLI. The tokenized credentials link to a user’s existing Mastercard, retaining rewards, credit lines and card benefits. Users and issuers can set purchase limits, merchant categories and permitted transaction locations. AgentCard is available at agentcard.ai.
→ Action
Payments authentication: Check passkey, customer-authorization and spending-rule controls for agent-initiated card payments.
Read article → from PR Newswire
House votes to modernize bank oversight and kill the penny
For the regulatory-affairs head at a regional bank
Your exam-readiness program now faces a second audience: Congress. Decide whether your AI, bank-run and cyber controls can withstand questions about whether regulators have the tools to test them.
The House passed the FUTURES Act, H.R. 8278, by a 417-7 vote this week. The bill would require financial regulators to assess their technology capabilities and identify ways to make examinations more efficient, effective and responsive. Reps. Marlin Stutzman, R-Ind., and Bill Foster, D-Ill., sponsored the measure. Foster said agentic AI could intensify bank runs and cybersecurity risks, and that financial agencies must continuously assess their technology so Congress can better protect consumers.
→ Action
Regulatory affairs: Inventory AI, bank-run and cybersecurity controls and the evidence available for regulatory examination.
Read article → from PYMNTS
GenWay taps Tavant AI to cut mortgage cycle times
For the mortgage operations head at a regional bank
GenWay sets a direct test: can your team cut file time while proving that automation handles non-QM and government-loan exceptions without weaker controls? If not, faster lenders may gain an edge in wholesale and correspondent channels.
GenWay Home Mortgage has put Tavant’s AI-powered TOUCHLESS platform into live use across its non-QM and Ginnie Mae divisions. Within three months, GenWay deployed six underwriting automation products for data and document processing, workflow automation and guided task execution. The tools are configured for non-QM and government-lending compliance and documentation requirements, and GenWay aims to scale without proportional headcount growth. Tavant says lenders using TOUCHLESS for agency and non-agency products have cut operating costs by as much as 60%, raised underwriting throughput four to 12 times, and reduced cycle times from 30 to 45 days to seven to 15 days.
→ Action
Mortgage operations: Test document-extraction exception queues and audit trails against non-QM and government-loan file requirements.
Read article → from FinTech Global
General AI
Large language models & AI infrastructure
Cohere encrypts Model Vault AI inference
For the CISO of a regional bank
Hosted-model due diligence can no longer stop at encryption in transit and at rest. Require proof that prompts, model processing and responses stay outside vendor and cloud-provider access during inference.
Cohere added confidential computing to the Encrypted tier of its single-tenant Model Vault inference platform. The service uses a confidential VM with Intel TDX or AMD SEV-SNP on CPUs and Nvidia GPUs in confidential computing mode. Cohere says customer data remains encrypted through memory, CPU-GPU interconnects and other hardware components, and is decrypted only within the protected execution environment. Each inference returns an attestation report for customers to verify the hardware, software and security policies. The feature is available now at no added cost to Model Vault customers; Cohere plans to open-source the full serving stack for independent audit.
→ Action
Third-party risk: Require inference attestation and verify whether hosted-model prompts remain readable in CPU and GPU memory.
Read article → from Venturebeat
NVIDIA Vera Rubin NVL72 delivers up to 3.7x GB300 throughput
For the chief technology officer at a regional bank
Do not lock a self-hosted inference refresh to GB300 economics before testing Vera Rubin. NVIDIA’s claimed 2.5x to 3.7x throughput gain could cut the racks needed for bank workloads, but it also increases vendor concentration and deployment-timing risk.
NVIDIA submitted preview Vera Rubin NVL72 results to MLPerf Inference v6.1, reporting up to 3.7x higher throughput than GB300 NVL72. On Qwen3-VL, the result covered offline, server and interactive scenarios using vLLM with NVIDIA Dynamo. On DeepSeek-R1, Vera Rubin NVL72 delivered up to 2.5x higher throughput than GB300 NVL72 using NVIDIA TensorRT-LLM. NVIDIA also reported that a four-rack, 288-GPU GB300 NVL72 DeepSeek-R1 setup reached 99% scaling efficiency in the offline scenario. NVIDIA says post-submission software results have not yet been verified by MLCommons.
→ Action
Infrastructure procurement: Reprice the next GPU cluster using 2.5x and 3.7x throughput scenarios, and require workload tests before committing to GB300.
Read article → from Nvidia
OpenAI introduces Astra for Law
For the general counsel of a regional bank
Astra gives lawyers enough research and workflow support to make informal adoption likely. Set approved matters, source checks and confidentiality rules before legal teams put confidential work into it.
OpenAI introduced Astra for Law, a GPT-6 Astra configuration for law firms and legal technology companies. It pairs legal-analysis instructions with a U.S. search index covering case law, statutes, regulations, court rules and administrative decisions from more than 230 million URLs, with sources added daily. At the highest reasoning effort, it passed 54.0% of overall-correctness checks on 200 Vals AI legal research questions, versus 38.7% for GPT-6 Astra using web search alone. Selected law firms will get Trusted Access in ChatGPT and Codex; eligible firms receive API Zero Data Retention, while ChatGPT Enterprise use is excluded from human review by default. OpenAI also launched 26 plugins connecting ChatGPT to tools including Relativity, Clio and iManage.
→ Action
Legal operations: Update legal-use guidelines to require approval before confidential bank matters enter Astra or connected legal-tool workflows.
Read article → from OpenAI
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime