BankingNewsAI Daily Brief ·
U.S. bank examiners now require AI governance controls like kill switches and vendor risk.
Banking AI
Financial institutions & fintech technology
U.S. bank exams now treat AI governance as routine supervisory scope (kill switches, vendor risk, lending controls)
U.S. banking regulators have escalated AI oversight by embedding AI use and governance into standard bank examinations, with specific attention to lending/underwriting impacts, third‑party model risk, and “kill switch”/human override controls. This is a shift from episodic model-risk reviews to standing exam expectations that will surface gaps faster—especially where generative AI is being used outside traditional SR 11-7 style controls.
Action
Accelerate a “exam-ready” AI control pack: inventory AI use cases (including vendor tools), document override/kill-switch procedures, and harden lending/underwriting monitoring for bias and drift before your next exam cycle forces an ad hoc scramble.
FSB opened consultation on responsible AI practices—global baseline is forming faster than country-by-country rules
The Financial Stability Board has opened a consultation on sound practices for responsible AI adoption in financial services. Even though it’s not a single binding rulebook, it’s a strong signal of converging expectations on governance, accountability, model risk management, and third-party dependencies across major jurisdictions.
Action
Align your AI risk framework to the emerging FSB baseline (governance, testing, monitoring, third-party controls) to reduce rework across regions and to pre-empt supervisory findings when local regulators map to FSB language.
General AI
Large language models & AI infrastructure
Anthropic shipped Claude Fable 5 broadly—and it’s already embedded across major enterprise/dev platforms with a new cost/limits regime
Anthropic released Claude Fable 5 for general availability (with Claude Mythos 5 restricted), positioning Fable 5 as its first broadly available “Mythos-class” model and claiming top performance on most benchmarks. It launched already integrated in tools that will show up inside your vendor stack (Microsoft Foundry, GitHub Copilot App/CLI, Notion, Cursor, Devin, Replit, etc.), but access is capacity-managed: included in some plans only until June 22, then credit-based, plus rate-limit resets after heavy demand. Net: capability is jumping, but procurement and FinOps will feel it immediately via token-hungry behavior and shifting entitlements.
Action
Interrogate your key AI-enabled vendors (Microsoft/GitHub, Notion, any dev/ops copilots) on whether Fable 5 is now in their stack, then demand a written view on (1) pricing pass-through, (2) rate-limit/credit throttles, and (3) how they route or fail over when capacity is constrained.
OpenAI faces a multi-state attorneys general investigation—enterprise AI procurement is about to get more legalistic
TechCrunch reports OpenAI is being investigated by state attorneys general, with questions spanning ad policies and handling of health data. The immediate change is not model capability—it’s the compliance surface area and due diligence burden likely to expand for anyone standardizing on a major model provider.
Action
Tighten your vendor governance: require clearer data-retention, training-use, and audit/right-to-notify clauses in LLM contracts, and be prepared for faster policy changes that could break existing approvals.
KPMG pulled an AI adoption report after hallucinations (including false claims about UBS)—trust in AI-generated “market intel” is now a risk item
KPMG withdrew an AI adoption report after apparent AI-generated inaccuracies, including incorrect statements about UBS. This is a concrete reminder that “AI-written research” can introduce reputational and decision risk even when it comes from brand-name institutions.
Action
Mandate provenance controls for external AI-generated research used in strategy/board materials (source checking, citation requirements, and an internal red-team review) to avoid making commitments based on fabricated claims.