BankingNewsAI Daily Brief ·
OpenAI releases GPT-6 Astra across ChatGPT Work, Codex and its API
Banking AI
Financial institutions & fintech technology
Half of BNY employees are building their own AI agents
For a payments operations executive at a regional bank
BNY’s results make employee-built agents a payments control issue, not a staff productivity test. Your bank must decide who may build agents that touch payment checks, and how it will prove their decisions are accurate and governed.
BNY’s Global Payments and Trade team is deploying digital employees to validate payments, read vendor addresses on cross-border transactions, and verify API and country codes. BNY defines digital employees as agentic systems that autonomously perform tasks alongside humans. BNY says the work fell from five or six minutes when done by a person to under 30 seconds with agents, while open payment investigations dropped by nearly 80%. Every BNY employee is trained on its enterprise AI platform, Eliza, which BNY says follows its data, risk, legal and compliance standards. BNY says 60% of employees use AI daily and 50% build their own agents.
→ Action
Payments operations: Inventory agent use in payment validation and require testing, owner approval and exception review before staff-built agents enter production.
Read article → from Arc-codex
Friday Harbor moves Freddie Mac income checks before underwriting
For the mortgage production head at a regional bank
Your bank should treat earlier agency income output as a production-control tool, not a transfer of credit risk. It must decide whether its loan-origination workflow can preserve the certificate and evidence behind the calculation.
Friday Harbor integrated Freddie Mac’s Income Calculator into its AI pre-underwriting platform for wage earners with variable pay. Lenders can send digitized paystub and W-2 data through an API, and Freddie Mac’s tool calculates monthly qualifying income before formal underwriting. Friday Harbor flags documentation or guideline issues, updates conditions as documents arrive, and ties findings to Freddie Mac’s Single-Family Seller/Servicer Guide. The calculated income can go through Loan Product Advisor, while the Income Calculator Certificate stays in the loan file for applicable income-calculation representation-and-warranty relief. The certificate does not remove lender responsibility for the loan’s accuracy or eligibility and does not provide blanket protection from repurchase demands.
→ Action
Mortgage operations: Validate document-transfer, certificate-retention and exception-handling controls before using Income Calculator results in borrower discussions.
Read article → from Nationalmortgageprofessional
Castellum.AI agent platform cuts AML alert review 83%
For a BSA/AML executive at a regional bank
Autonomous alert closure turns a staffing tool into a model-governance decision. Do not accept vendor closure rates until independent testing proves false negatives, escalation thresholds and audit records work on your data.
Castellum.AI launched Arbiter AI, which integrates with LexisNexis Bridger Insight XG to resolve sanctions, PEP and adverse-media AML alerts. The agent gathers supporting data, applies institution-configured risk logic and creates an audit trail before deciding whether to close an alert without human review. Castellum.AI says Arbiter can resolve up to 95% of alerts and cut review time by 83%, based on its own testing rather than independent benchmarking. The company says its decisions are calibrated to each institution’s risk thresholds and documented with adjudication steps. EU AI Act requirements for high-risk systems, including validation, oversight and explainable decision logic, take effect December 2, 2027.
→ Action
Model risk: Test alert closures for false negatives and traceable audit records using your screening data.
Read article → from Autonainews
General AI
Large language models & AI infrastructure
OpenAI makes GPT-6 Astra available in ChatGPT Work, Codex and API
For the technology chief at a regional bank
API governance does not limit automation: agents can use applications without APIs. You need to decide which systems they may enter and where approvals and audit checks must stop them.
OpenAI made GPT-6 Astra available in ChatGPT Work, Codex and its API. In ChatGPT Work and Codex, Astra can write code and use the applications employees use, including applications without an API. OpenAI says enterprise administrators can limit Astra to approved websites and desktop applications, manage uploads and downloads, and control browsing history. Confirmation policies can require approval before consequential actions, while automated review can assess potentially unsafe or unauthorized tool calls. Astra produced unintended outcomes 89% less often than GPT-5.6 Sol in OpenAI’s internal computer-use safety benchmark.
→ Action
Identity and access: Inventory desktop applications reachable by computer-use agents; require approvals for data sharing, data deletion, uploads and downloads.
Read article → from OpenAI
Anthropic CEO urges AI companies to slow model development
For the chief information-security officer at a US bank
Your AI vendors may be advancing faster than their controls can prove safe. Require evidence that agent permissions, access limits and shutdown controls work before expanding autonomous use.
Anthropic CEO Dario Amodei urged AI companies to slow model development. Amodei addressed the AI Impact Summit in New Delhi, India, on February 19, 2026.
→ Action
Third-party risk: Require AI providers to document agent access limits, approval rules, audit logs and emergency shutdown controls before production use.
Read article → from Reuters
DOJ probes Nvidia’s $20 billion Groq deal
For the technology procurement head at a large bank
Do not treat Nvidia’s Groq deal as a durable source of inference capacity. Long-term AI contracts need chip and rack alternatives that can work with Nvidia infrastructure if Groq’s role changes.
The US Department of Justice has launched an antitrust probe into Nvidia’s $20 billion deal to license Groq’s AI accelerator technology and hire key engineering staff. Groq’s inference-as-a-service business remains intact, but Nvidia obtained mature silicon and the engineers needed to develop it. Nvidia unveiled LPX racks at GTC in March that use 256 Groq-3 accelerators, which are described as lightly modified Groq-2 designs. Nvidia says Groq-3 combined with Vera Rubin GPU racks will serve inference workloads. Cerebras works with AWS and AMD, SambaNova works with Intel, and d-Matrix combines its platform with Nvidia GPUs. Nvidia also opened MGX rack designs and offers NVLink Fusion licenses.
→ Action
AI infrastructure procurement: Require exit rights and compatible alternative-chip options in long-term inference capacity contracts.
Read article → from Theregister
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime