How Bank Regulators Are Treating Frontier AI

Federal financial regulators are not writing frontier-AI rules from scratch — they are narrowing an existing supervisory document and standing up a monitoring body. In May 2026, the Federal Reserve, together with the OCC and FDIC, amended its 2011 model-risk-management guidance (SR 11-7) to say it does not apply to generative or agentic AI, while the Financial Stability Oversight Council (FSOC) created a standing interagency AI Working Group rather than issue substantive AI rules of its own. Neither move is a comprehensive AI rulebook for banks. Both are real, citable, and distinct from the executive-order- and NIST-level policy content already on CASRAI — this is what the sector-specific, prudential-regulator layer actually looks like so far. CASRAI’s own NIKOLAI reference tracks a closely related concept — dual-use capability risk — that shows up directly in how one of these regulators is talking about a named frontier model.

Key facts

Guidance narrowed: SR 11-7, the Fed/OCC/FDIC’s joint 2011 model-risk-management guidance, was amended to clarify it does not apply to generative or agentic AI.
Who amended it: The Federal Reserve, “together with the OCC and FDIC,” per Vice Chair for Supervision Michelle Bowman.
Announced: May 1, 2026, in a speech at an FSOC AI Series Roundtable on Cybersecurity and Risk Management.
Underlying letter: The amendment traveled under SR letter 26-2, dated April 17, 2026.
FSOC’s structural response: A standing interagency AI Working Group, created via the Council’s 2025 Annual Report (cover-dated November 26, 2025), Section 3.4.
Legal anchor for “AI”: FSOC’s definition tracks Executive Order 14179 and 15 U.S.C. 9401(3), the same statutory AI definition used across the federal AI Action Plan.
OCC’s own position: Not established either way as of this guide’s publication — see the caveat below.

What the Fed, OCC and FDIC Actually Changed

SR 11-7, formally the Supervisory Guidance on Model Risk Management, is the joint Federal Reserve/OCC document banks have used since 2011 to govern how they validate, document, and monitor internal quantitative models — credit scoring, stress testing, valuation, and the like. It long predates generative and agentic AI, and examiners and banks alike have spent the last few years arguing over whether, and how, a framework built for static statistical models should apply to a large language model or an autonomous agent.

Bowman’s May 1, 2026 answer, delivered at an FSOC AI Series Roundtable on Cybersecurity and Risk Management, was that it largely shouldn’t — not because generative and agentic AI are risk-free, but because SR 11-7’s specific validation machinery doesn’t fit them. In her words: “Together with the OCC and FDIC, the Fed recently amended our model risk management guidance to clarify that it does not apply to generative or agentic AI.” The amendment itself traveled as SR letter 26-2, dated April 17, 2026. Bowman also flagged a second, related change in progress: the agencies’ third-party risk management guidance is being updated “to eliminate vagueness and avoid hindering innovation,” a acknowledgment that the existing vendor-risk framework was creating friction for banks adopting AI tools through third parties rather than building them in-house.

What this is not is a statement that generative and agentic AI fall outside supervision. Bowman was explicit that the Fed’s approach stays centered on “identifying and addressing material financial risks that could lead to bank failures,” and that examiners still need to understand specific use cases, material impacts, and consumer effects before signing off on an AI deployment. The change is scoping, not deregulation: SR 11-7’s particular tools no longer apply, but the underlying supervisory expectation — understand what you built, monitor it, be able to explain it — still does. A bank that reads the SR 11-7 carve-out as “AI is unregulated” would be reading past what the Vice Chair actually said.

Bowman Named a Real, Dual-Use Frontier Capability as the Example

The same speech is notable for something more specific than a policy update: Bowman named an actual frontier AI capability as her illustration of why this space is moving fast. “Anthropic’s Mythos—an AI model that identifies cyber vulnerabilities—highlights the dynamic nature of this technology and the rapid pace that its capability can evolve,” she said, describing high-level coordination between Treasury and Federal Reserve leadership in response.

Read that framing carefully, because it’s doing specific work: Bowman is describing Mythos as dual-use in exactly the classic sense — a capability that is valuable defensively (a bank or a security vendor uses it to find and patch its own vulnerabilities before an attacker does) and also creates offensive risk (the same vulnerability-discovery capability, misused, finds vulnerabilities for someone other than the defender). That is a supervisor naming a specific, real capability and using it to illustrate a category of risk, not a comprehensive Fed position on Anthropic or on vulnerability-research AI generally. This guide reports that framing as Bowman’s and the Fed’s own characterization — not CASRAI’s assessment of Mythos, and not Anthropic’s own description of the tool.

FSOC’s Own Move: A Working Group, Not a Rulebook

While Bowman’s speech addressed a specific supervisory document, FSOC’s 2025 Annual Report (cover-dated November 26, 2025) took the broader, structural step. Section 3.4, “Harnessing Artificial Intelligence to Promote Financial Stability,” recommends “that member agencies use its Artificial Intelligence Working Group to explore opportunities for AI to promote the resilience of the financial system.” The working group’s mandate, per the report, covers three things: monitoring “potential risks to financial stability that might be posed by the adoption of AI both within and outside the financial services sector,” identifying high-value AI use cases member agencies can adapt for their own supervision and regulation, and providing “a forum for public-private dialogue to identify regulatory impediments” to responsible AI adoption.

The report anchors its AI definition explicitly to Executive Order 14179 and its underlying statutory text, 15 U.S.C. 9401(3): AI as “a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments.” The same section also ties into the federal government’s internal AI posture — noting that America’s AI Action Plan recommends “mandating appropriate training and access for employees whose work could benefit from access to frontier language models,” governed by OMB Memoranda M-25-21 and M-25-22, and records that Treasury appointed its own Chief AI Officer in June 2025.

What Section 3.4 does not do is set binding AI-specific capital, disclosure, or model-validation requirements for financial institutions. It is a monitoring-and-coordination mandate, standing up a body and a mission rather than a rule. Read alongside Bowman’s SR 11-7 announcement, the two moves are complementary: the working group is the interagency structure for watching AI-driven financial-stability risk broadly, while the SR 11-7 amendment is the Fed’s own concrete supervisory-scoping decision inside that broader watch.

What We Could Not Confirm: The OCC’s Own Risk Report

One natural question is whether the OCC — as national-bank supervisor and a co-signer of the SR 11-7 amendment — has published its own AI-specific risk guidance separately from the joint Fed action. We checked the OCC’s Fall 2025 Semiannual Risk Perspective landing page, which lists the report (published December 16, 2025) but whose visible summary content contains no AI-specific risk section. We did not review the full PDF of that report. That means we can say the OCC’s own summary page doesn’t surface AI as a named risk category — but we cannot say, and do not claim, that the OCC has no AI-specific supervisory guidance elsewhere in that document or in other publications. That question remains open pending a direct review of the full report, and this guide will be revised once that review happens.

How This Differs From the EO/NIST Layer Already on CASRAI

CASRAI already covers the general federal AI-policy layer in depth: EO 14179 and EO 14365 for the executive-order framework, and NIST’s Govern-Map-Measure-Manage functions for the voluntary risk-management standard most agencies point to. What that existing content doesn’t cover is the prudential-regulator layer specifically — the bank, thrift, and credit-union supervisors who examine individual institutions rather than set economy-wide policy. That’s the gap this guide fills: it’s a sector-specific companion to the EO/NIST material, not a restatement of it. FSOC’s own citation of EO 14179 (above) is the connective tissue between the two layers — the financial regulators are applying the same statutory AI definition the executive-order track already established, not inventing a separate one.

Where NIKOLAI Fits

Bowman’s Mythos framing — a named capability that is useful for defense and exploitable for offense — is an informal, real-world instance of a concept CASRAI’s own NIKOLAI dictionary tracks formally as Risk Domain, one of the 64 elements in NIKOLAI’s independent, unendorsed reference of frontier-AI-safety terminology. Risk Domain is the top-level category of catastrophic harm a safety framework scopes itself to — a controlled list that, across the frontier labs’ own published frameworks (Anthropic, OpenAI, Google DeepMind, and others) and the EU’s GPAI Code of Practice, consistently includes CBRN, Cyber offence, loss of control, and harmful manipulation as parallel categories.

Mythos, as Bowman describes it — a model that identifies cyber vulnerabilities, useful for defenders and misusable by attackers — is a textbook example of what “Cyber offence” as a risk domain is meant to capture: the same underlying capability, defensive in one hand and offensive in the other. To be precise about what this is and isn’t: NIKOLAI’s Risk Domain crosswalk maps the frontier labs’ own published frameworks and a handful of regulators with scoped AI-safety frameworks (the EU AI Act, California’s SB 53); neither Bowman’s speech nor the SR 11-7 amendment appears as a mapped row in that crosswalk, and none should be inferred from this guide. This section is CASRAI’s own explanatory tie-in, illustrating a concept the Fed is gesturing at informally with a term NIKOLAI defines formally — not a claim that the Fed relied on, referenced, or endorses NIKOLAI in any way.

For the fuller picture of how NIKOLAI organizes frontier-AI-safety terminology, see the map of all ten NIKOLAI tracks. For how frontier labs’ own dual-use safety commitments compare, including Anthropic’s, see which frontier labs let outsiders audit RSP compliance and the broader frontier AI labs overview, which covers Anthropic’s Responsible Scaling Policy directly.

FAQ

Does SR 11-7 still apply to any AI a bank uses?

Yes, for AI that fits the traditional model-risk profile SR 11-7 was written for — conventional statistical and machine-learning models used for things like credit scoring or stress testing. The May 2026 amendment narrows SR 11-7’s scope to exclude generative and agentic AI specifically; it does not repeal SR 11-7 for everything else.

Does the SR 11-7 change mean generative and agentic AI are unregulated at banks?

No. Bowman was explicit that supervision still focuses on material financial risk, specific use cases, and consumer impact. What changed is which guidance document’s specific validation requirements apply — not whether banks’ use of generative or agentic AI is subject to supervision at all.

What is Mythos?

Per Vice Chair Bowman’s May 1, 2026 speech, Mythos is “an AI model that identifies cyber vulnerabilities,” which she attributes to Anthropic. This guide reports that characterization as Bowman’s and the Fed’s own framing of the model as an example of fast-evolving, dual-use AI capability — not an independent CASRAI or Anthropic description of the tool.

Has the OCC issued its own AI-specific risk guidance?

Unconfirmed either way. The OCC’s Fall 2025 Semiannual Risk Perspective landing page doesn’t display an AI-specific section in its summary, but the full report was not reviewed for this guide, so an absence cannot be confirmed. Treat this as an open question, not a finding.

Is NIKOLAI a standard the Fed or FSOC uses?

No. NIKOLAI is CASRAI’s own independent, unendorsed reference dictionary. Nothing in Bowman’s speech or the FSOC Annual Report references NIKOLAI, and this guide does not claim otherwise — the NIKOLAI section above is CASRAI’s own explanatory tie-in to an existing regulatory statement, not evidence of any relationship between the two.