BankingNewsAI Daily Brief ·
JPMorgan Payments develops agentic commerce tools for merchants
Banking AI
Financial institutions & fintech technology
JPMorgan Payments to develop agentic commerce for merchants
For the head of merchant payments
Merchant acquiring now needs controls for who an AI agent is, what a customer approved and how much it can spend. Waiting for industry standards leaves your merchants exposed as LLM channels begin to take transactions.
JPMorgan Payments plans to develop agentic-commerce capabilities for merchants with technology firm Mirakl, with a launch due later this year. The service will let merchants sell through LLM channels including Gemini, Copilot and Perplexity. JPMorgan and Mirakl aim to build infrastructure for autonomous AI-agent shopping while maintaining consumer and business security and control. JPMorgan plans to integrate with consumer agents, manage agentic fraud risk and support agent-driven channels. It said merchants of all sizes can participate without building or managing complex integrations and protocols, using an AI commerce engine and payment infrastructure for secure agent transactions.
→ Action
Payments authentication: Check customer-permission, agent-identity, spending-limit and fraud controls for agent-initiated merchant payments.
Read article → from Retail-systems
Chatham scales capital markets expertise with OpenAI
For a capital-markets operations leader at a regional bank
You need to decide which trade controls can move from manual evidence gathering to AI comparison without moving approval authority. The standard is not speed alone: it is a review trail that lets experienced staff test exceptions and retain judgment.
Chatham Financial uses Codex and GPT-5.6 to build tools and redesign workflows, reducing trade validation from about 30 minutes to under four minutes. Its trade validation application gathers transaction evidence, compares key terms, and flags discrepancies for review. Chatham is comparing application results with experienced reviewers on real transactions before expanding automation. The firm plans to extend the application to more trade types and automate more of the workflow while retaining controls and professional oversight.
→ Action
Capital markets operations: Map trade-validation evidence sources, comparison rules, and escalation points; test AI outputs against experienced reviewers before automating approvals.
Read article → from OpenAI
SEC charges entities in $15 million WhatsApp investment fraud schemes
For the chief fraud officer at a regional bank
Your fraud controls must treat AI investment claims, WhatsApp recruiting and SEC credentials as a linked scam pattern, not separate warning signs. Payment intervention needs to catch the customer before funds reach a fake platform.
The SEC charged Cryptoaiml Ltd., Cryptoaiml Capital Foundation, TSAI Pro Ltd. and TSAI Capital Foundation over schemes that allegedly took more than $15 million from hundreds of retail investors, including many in the U.S. Cryptoaiml allegedly used WhatsApp chats, impersonated investment professionals and claimed AI-generated trading tips produced large profits. Investors transferred crypto assets to a fake platform, where no trading took place and displayed profits were fictitious. TSAI allegedly offered an AI trading bot program through its website, WhatsApp and Facebook, promised guaranteed profits and said investors could earn money by recruiting others. Both groups allegedly claimed SEC regulation and posted falsified Form D records or agency credentials.
→ Action
Fraud operations: Add alerts for customers sending crypto after WhatsApp investment contacts and verify SEC-registration claims in intervention scripts.
Read article → from SEC
General AI
Large language models & AI infrastructure
Microsoft releases 2026 Digital Defense Report
For the CISO of a regional bank
Treat each AI agent as an identity with its own permissions, tool access and audit trail. Your existing identity controls are the starting point, but they must cover agent-to-agent authentication, revocation and activity across systems.
Microsoft released its 2026 Digital Defense Report, which says threat actors use AI for reconnaissance, social engineering, malware and exploit development, and post-compromise activity. Microsoft says AI can make social-engineering campaigns more targeted and automate parts of attack workflows. The report says AI agents can access enterprise data, applications, APIs and tools at different levels of autonomy. The report looks at agent identity, access, authentication, attribution and revocation, as well as defenses for prompt injection, memory, models, data and agent behavior. Security teams work with information from endpoints, identities, cloud environments, applications, email and networks.
→ Action
Identity and access management: Inventory AI agents, map each agent’s data and tool permissions, and test authentication, logging and access revocation.
Read article → from Microsoft
AWS releases open-source Strands Decider 2B
For the chief technology officer at a regional bank
Do not treat an open, low-latency decider as a control until your bank can prove its confidence scores hold on its own agent cases. AWS supplies a model judgment, but your teams still set thresholds and decide whether an action proceeds.
AWS released Strands Decider 2B, an open model that selects from bounded choices and returns probabilities for agent workflows. The roughly 2-billion-parameter model is fine-tuned from Alibaba Qwen and is free under an Apache 2.0 license, but users must provide their own computing resources. AWS says it can route requests, select tools, evaluate outputs, or review proposed agent actions; application code can then proceed, deny, request confirmation, or send feedback. On 230 local RTX 3090 requests, version 18 had 106 milliseconds median latency and 296 milliseconds at the 95th percentile. Version 19 scored about 72% accuracy and a 0.35 Brier score on public JevBench data.
→ Action
Model risk: Test calibration and error rates on bank tool-call cases before using confidence thresholds to approve or deny agent actions.
Read article → from Venturebeat
NVIDIA offers DGX Spark 64GB from $4,999
For the infrastructure chief at a regional bank
NVIDIA gives your bank a cheaper path to run sensitive agent work outside public-cloud inference. You now need to decide which workloads justify local hardware, rather than assume cloud is the default for every model.
NVIDIA will offer a 64GB DGX Spark configuration through Acer, ASUS, Dell, Gigabyte, HP and MSI starting Oct. 23 at $4,999. NVIDIA says one system can run models of up to 100 billion parameters on device, with DGX OS and its AI software stack included. The system supports agent development, inference, fine-tuning, data science and edge development, and includes NVIDIA ConnectX-7 networking. Two units can connect by QSFP cable, pool 128GB of memory and support models of up to 200 billion parameters. NVIDIA says its Qwen 3.8 27B test produced up to 1.7x the performance of one system.
→ Action
AI infrastructure: Inventory agent workloads that use sensitive data and compare local DGX Spark capacity, hardware cost and cloud-inference spend.
Read article → from Nvidia
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime