BankingNewsAI Daily Brief · Saturday, April 25, 2026
Banking AI
Financial institutions & fintech technology
US banking regulators replaced SR 11-7: model risk governance is being reset for modern AI
The OCC, Federal Reserve, and FDIC issued updated interagency Model Risk Management guidance (SR 26-2), formally rescinding SR 11-7. This matters because it refreshes examiner expectations for how banks inventory, validate, monitor, and govern models—now including newer ML/AI and more complex third-party/model supply-chain realities.
Action
Re-baseline your model inventory and validation program against SR 26-2 (especially for GenAI/agent workflows and vendor models) and pre-wire evidence artifacts examiners will ask for: intended use, limitations, monitoring, change control, and accountability.
Commonwealth Bank deployed an agentic AI system that detects new scam patterns and generates new defenses
Commonwealth Bank says it has deployed an agentic AI system to spot emerging fraud/scam patterns in transactions and payments data and help generate the rules used to intercept them. The key shift is from static detection models to an automated detect→propose-controls loop designed to keep pace with rapidly mutating scam tactics.
Action
Stand up an equivalent ‘fraud rules factory’ capability: combine real-time anomaly detection with governed, human-approved rule generation and rapid rollout; measure impact in time-to-mitigate and false-positive cost, not just model AUC.
General AI
Large language models & AI infrastructure
DeepSeek previewed V4: longer-context, more efficient open model that narrows the frontier gap
DeepSeek released a preview of its V4 flagship model, highlighting much longer prompt handling and efficiency improvements, and keeping its open approach. The notable change is that high-capability, long-context models are increasingly available outside the closed frontier labs—raising both competitive pressure and data-governance complexity.
Action
Expand your model risk and procurement playbook to include high-performing open models: add red-teaming, secure fine-tuning patterns, and licensing/compliance checks so teams don’t adopt ‘shadow’ alternatives when closed-model cost or policy constraints bite.