BankingNewsAI Daily Brief ·
European Commission publishes guidelines on prohibited AI practices under the AI Act
Banking AI
Financial institutions & fintech technology
Sokin launches MCP connector for AI payment preparation
For the payments head at a mid-size bank
AI can now prepare beneficiaries, FX exchanges and payment instructions from live account data. Your controls must work before settlement, not only at the approval screen, or faster instruction creation will widen the error and fraud window.
Sokin released a Model Context Protocol connector that links its licensed payment infrastructure to customers’ AI tools. Finance teams can use ChatGPT, Claude, Gemini or Perplexity to retrieve live balances, FX quotes and payment status, and prepare cross-border payments for approval. The connector can create a beneficiary, prepare a currency exchange between supported wallets, and prepare a payment. Customers must confirm every instruction, and configured Sokin approval rules apply before execution. Sokin holds licenses in 36 countries and moves money to more than 170 countries in more than 70 currencies.
→ Action
Payments authentication: Check beneficiary-creation, customer-confirmation and approval controls for AI-prepared payment instructions.
Read article → from PR Newswire
Visa adds Featurespace AI to flag fraud before money leaves an account
For the fraud chief at a regional bank
Visa is selling broader fraud insight before transfer authorization, but its claims lack detection and false-alert results. Your team must decide whether faster scam interdiction justifies testing a score whose operational burden is unproven.
Visa enhanced A2A Protect with a unified fraud score that combines its risk intelligence with Featurespace technology before funds leave a customer account. Visa says this is its first in-market integration of Featurespace technology. The service uses AI and transfer learning to give banks global A2A risk insights without waiting for models to learn from only their own transaction data or for a data-sharing consortium to form. Banks can opt into network-level intelligence on scam hotspots and coordinated fraud activity. Each alert includes a plain-language explanation, and Visa says banks can connect through a single API to existing payment and fraud-management processes. Visa provided no detection or false-alert reduction figures.
→ Action
Fraud operations: Test score thresholds, alert explanations and manual-review capacity against account-to-account payment flows before deployment.
Read article → from Kachwanya
General AI
Large language models & AI infrastructure
Anthropic introduces Claude Fable 5.1 and Mythos 5.1
For the CIO of a large US bank
Fable’s pricing and customer-controlled storage remove two objections to high-volume agent use, but EFS will arrive in phases. Decide whether eligible zero-data-retention access is enough now or whether your most sensitive workloads must wait.
Anthropic released Claude Fable 5.1 for general availability and made Claude Mythos 5.1 available through trusted access programs. The models use the same underlying model but have different safeguards; Mythos supports cybersecurity and life-sciences work. Anthropic says Fable 5.1 costs about 25% less than Fable 5 for typical token-billed workloads because cache-read prices are lower, with savings up to about 45% for highly agentic work. Its Enterprise Frontier Safeguards store data in cloud infrastructure controlled entirely by the customer rather than Anthropic, and will reach enterprise customers in phases beginning later this fall. Eligible customers can use Fable 5.1 with zero data retention until EFS is available.
→ Action
AI platform: Map agent workflows with repeated context and quantify cache-read use; verify eligibility for zero data retention before EFS.
Read article → from Anthropic
Commission publishes AI Act prohibited-practices guidelines
For the AI compliance head at a US bank
Non-binding does not mean optional under the AI Act. You need a current inventory of AI uses that could be seen as manipulation, social scoring or real-time remote biometric identification before the Commission’s view becomes a dispute.
The European Commission published guidelines on AI practices prohibited under the AI Act. The guidance covers harmful manipulation, social scoring and real-time remote biometric identification, among other practices. It provides legal explanations and practical examples intended to help stakeholders comply and to support consistent application of the AI Act across the EU. The AI Act sorts AI systems into prohibited, high-risk and transparency-obligation categories. The guidelines are non-binding, and only the Court of Justice of the European Union can give authoritative interpretations of the prohibitions.
→ Action
AI governance: Inventory AI uses involving manipulation, social scoring or real-time remote biometric identification and map them to the guidance.
Read article → from Europa
Microsoft adds Grok to Word, Excel and PowerPoint
For the Microsoft 365 administrator at a bank
Your Office AI controls can no longer treat Copilot as an OpenAI-and-Anthropic service. Decide whether Frontier users may access Grok before employees select it in documents, spreadsheets and presentations.
Microsoft has added SpaceXAI’s Grok models to the Copilot model picker in Word, Excel and PowerPoint for Microsoft Frontier users. The release is a preview, which Microsoft says will gather feedback. Office Copilot users can now choose among OpenAI GPT models, Anthropic Claude models and SpaceXAI Grok models. Microsoft added Grok models to its Foundry service last year before extending access to Copilot.
→ Action
Model risk: Test Grok outputs and data handling in Word, Excel and PowerPoint before permitting Frontier users to select it.
Read article → from The Verge
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime