The Fed, OCC, and FDIC are now asking banks seven specific questions about AI agent controls in routine exams — while SR 26-2, issued in April, explicitly excludes agentic AI from its scope. The gap is already live.
Table of Contents #
On October 13, 2026, Reuters reported that the Federal Reserve, Office of the Comptroller of the Currency, and Federal Deposit Insurance Corporation are now asking banks — at all asset-size tiers — how they monitor, limit, and shut down AI agents. The agencies are not yet prescriptive. They declined to comment publicly. But the questions are live, they are appearing in routine supervisory visits, and they are not matched by any existing guidance.
SR 26-2, issued April 17, 2026, replaced SR 11-7 as the primary model risk management framework for US banks. It is more detailed on data governance, documentation requirements, and third-party model risk than its predecessor. It also contains one sentence that a bank’s model risk officer needs to read very carefully: generative AI and agentic AI models are described as “novel and rapidly evolving” and explicitly outside the guidance’s scope. The agencies plan a future Request for Information on AI in model risk management to address this. That RFI has not dropped.
In the meantime, examiners are asking the questions.
The gap is not academic. A bank that has deployed AI agents — in fraud detection, document processing, customer service routing, trade surveillance, or any other production workflow — currently has no regulatory framework that maps specifically to agent governance. SR 26-2 does not cover it. There is no interagency agent-specific guidance. The OCC has encouraged expanded agentic AI use “with guardrails and human oversight,” which is encouragement to deploy without a framework to govern the deployment. When the examiner arrives and asks how the bank would shut down an autonomous system, there is no compliance documentation to point to. There is only the answer the bank can produce in the room.
This is the situation. Seven questions, no framework, and examinations that have already started.
Regulatory & Compliance Angle #
SR 26-2 is the governing document for model risk management at federally supervised banks. It extends SR 11-7’s conceptual framework — model inventory, independent validation, ongoing monitoring, model risk tiering — and adds new requirements around model documentation for complex statistical and machine learning models, third-party model dependencies, and challenger model programs. What it does not do is address generative AI or AI agents in any operational detail. The OCC bulletin accompanying SR 26-2 states plainly that these model types are out of scope and that further guidance is coming.
The absence of agent-specific guidance is not an oversight. The agencies understand that agentic AI represents a genuinely different risk profile: an agent does not make a single decision on structured inputs and return a score. It takes sequences of actions, calls external tools, modifies data, and — in many implementations — can spawn subagents or delegate to other systems. The validation, monitoring, and control frameworks built for traditional statistical models and even for LLMs used in constrained classification tasks do not map cleanly to systems that can initiate wire transfers, file reports, or execute trades autonomously.
The RFI, when it comes, will likely create an initial framework for agent governance, inventory requirements, and validation standards. But the timeline for that RFI, public comment, interagency reconciliation, and final guidance issuance is realistically 18 to 36 months. In that window, any bank that is running AI agents in production — which, per multiple industry surveys, is the majority of large and mid-size institutions — is operating in a governance gap. The examination questions are already in the room. The guidance is not.
The European Central Bank is moving faster. The ECB asked banks to submit plans for defense against AI-enabled cyberattacks by October 31, 2026. That mandate is separate from the EU AI Act’s financial services obligations, which carry their own timeline. For banks operating across jurisdictions, the question of agent governance is arriving simultaneously from multiple regulatory directions with incompatible specificity levels.
What the Examiner Will Find #
According to the Reuters report on the agencies’ examination questions, supervisors are now asking seven things. Banks that have not prepared answers should treat each one as a gap assessment item, because the examiner is treating it exactly that way.
The first question — whether the bank uses Chinese-developed technology — is not primarily about trade policy. It is a provenance question. The concern, explicitly connected to a CrowdStrike finding involving a South Korean banking campaign, is that adversarial infrastructure may be embedded in AI tooling without institutional awareness. Vendors regularly deploy open-weight models — Qwen, Kimi, DeepSeek, and others developed by Chinese AI labs — inside enterprise software without naming them in contracts or data sheets. A bank that answers this question by saying it has no exposure may be incorrect without knowing it. The correct answer requires a vendor-by-vendor audit of what models are embedded in licensed AI products, where those models were developed, and where inference is running. Most banks cannot produce that answer today.
The second question — whether AI agents are used in cyber defense — requires the bank to have an accurate inventory of where autonomous or semi-autonomous AI systems are operating. This is harder than it sounds. Many agentic AI deployments in banks have arrived through vendor products, not through internal development programs. A bank’s enterprise license for a document processing suite, a trade compliance platform, or a fraud investigation tool may now include agentic capabilities that were not present when the contract was signed and were not flagged as material changes by the vendor. The bank’s formal model inventory — which under SR 26-2 is required to cover all models that inform decisions — may not include these systems at all.
Questions three through six probe the governance controls around any agents the bank identifies: what limits apply to the system, how its actions are monitored and approved, whether the bank can track and audit agent decisions, what safeguards constrain the agent to its authorized scope, and whether humans remain involved in key decisions. Each of these questions has a specific evidence requirement. Authority limits are documented in system configuration, not asserted. Approval records exist in platform logs, not in the model risk officer’s recollection. Decision logs contain inputs, active instructions, model version, action taken, and approving human identity — not just a record that a tool was called. Examiners will ask to see this documentation, not just to be told it exists.
The seventh question is the one most banks will struggle with most visibly: how would you shut down an autonomous system? The instinctive answer — “we have a kill switch in the application” — describes a control that is most likely to fail precisely when it is needed. An agent that is behaving unexpectedly, running in a failure state, or that has been compromised does not reliably execute its own shutdown logic. A credible answer to question seven describes revocation of the agent’s credentials and network egress from infrastructure the agent does not control, with evidence that the procedure has been tested and that the timing from trigger to confirmed shutdown has been measured. Most banks have not run that test.
The Governance Gap #
The structural gap here is not just about missing documentation. It is about a regulatory posture that creates maximum examination exposure at minimum guidance specificity.
SR 26-2 was released in April. It replaced SR 11-7 and is the primary document that any bank’s model risk program is aligned to. By explicitly excluding agentic AI from scope, SR 26-2 tells banks: “agents are different, guidance is coming, figure it out for now.” The OCC bulletin adds: “we encourage expanded use with guardrails.” Neither document provides a framework for what guardrails are required, how they should be validated, or how agent programs should be inventoried and governed. Banks that have interpreted this as permission to deploy without a specific agent governance framework have made a reasonable reading of the guidance — and they are now the banks that cannot answer the examination questions.
The Chinese AI model provenance question exposes a second governance gap that sits below the agent governance layer. Enterprise AI procurement has, for the last several years, treated vendor AI products as black boxes that the bank neither can nor should audit at the model level. The contract governance approach — relying on vendor representations and contractual data protection clauses — was adequate when the risk was primarily around data handling. It is not adequate when the question is what model is running inference on the bank’s customer data and transaction records, who developed that model, and what the model’s training data provenance is. Banks that have not built model provenance requirements into their vendor procurement and renewal processes are now discovering that gap in an examination room.
The temporal problem compounds both issues. The RFI has not been issued. When it is issued, public comment periods typically run 60 to 90 days. Interagency reconciliation and final guidance typically takes another 12 to 18 months after comment close. By the time binding agent governance guidance exists, several full examination cycles will have occurred. Banks that wait for the guidance before building their agent governance programs will have already been examined against standards that do not yet exist in writing — but that examiners are applying anyway based on their own assessment of what sound practice requires.
The ECB deadline creates additional pressure for institutions operating in both jurisdictions. An interim agent governance framework built to satisfy the Fed/OCC/FDIC examination questions will not automatically satisfy the ECB’s AI-enabled cyberattack defense requirements, which approach the problem from a cyber resilience angle rather than a model risk angle. Banks with transatlantic operations are managing two incompatible regulatory clocks simultaneously.
The SuperML Take #
The seven examination questions represent exactly the kind of regulatory pressure that forces institutions to build real governance infrastructure rather than compliance theater. That is not a criticism. The governance questions being asked — inventory, authority limits, decision logs, human checkpoints, provenance, and tested shutdown procedures — are the right questions. The problem is that banks deploying AI agents at scale have been doing so without a regulatory framework requiring that infrastructure to exist.
The immediate consequence is that model risk management teams are being asked to answer questions for systems that were never routed through model risk review in the first place. Many enterprise AI agent deployments arrived as vendor software features, not as internally developed models subject to new model review and validation. They do not appear in the model inventory. They do not have validation documentation. They do not have approved authority limits or documented approval chains. They have none of the artifacts that question three through seven require — not because the bank was negligent, but because SR 26-2 did not require them and the procurement process did not capture them.
The correct response is not to draft documentation retroactively and hope it satisfies an examiner. The correct response is to build the agent governance infrastructure that the questions imply is required, starting with the things that take time: the agent inventory (because you cannot govern what you cannot enumerate), the vendor model provenance audit (because the Chinese model question requires a real answer, not a denial), and the shutdown procedure test (because “we have a kill switch” is an assertion, not evidence). These three deliverables will not be complete by the next examination for most banks. Banks that start now will be in a materially better position than banks that wait for the guidance.
The SR 26-2 gap is temporary. The agencies will issue the RFI, gather comments, and eventually publish agent-specific guidance. When that guidance arrives, it will almost certainly require exactly what the examination questions are already asking for: inventory, controls documentation, validation evidence, and tested shutdown procedures. Banks that build the governance infrastructure now — because the examination questions demand it, even absent the formal guidance — will have a head start on compliance with whatever the binding standard ultimately says. Banks that wait will be building under time pressure after the guidance drops.
The Chinese model provenance question is the one most likely to surface unexpected exposure. Every bank that has deployed enterprise AI products since 2024 should assume, until they have verified otherwise, that at least one of those products runs inference on a model developed by a Chinese AI lab. That assumption should trigger a vendor audit, not a dismissal. The audit will either confirm the assumption — in which case the bank needs to assess the risk and decide whether the deployment is appropriate — or refute it, in which case the bank has the documentation it needs to answer the examination question. Either outcome is better than discovering the answer for the first time in a supervisory conversation.
Sources #
- Reuters, “U.S. bank regulators scrutinize AI use in routine bank exams,” October 13, 2026: https://qz.com/us-bank-regulators-ai-scrutiny-routine-examinations-061226
- Beri.net analysis of Fed/OCC/FDIC supervisory questions: https://www.beri.net/article/fed-occ-fdic-bank-ai-agent-supervisory-questions-shutdown-audit-trail-chinese-models-sr-26-2-gap
- AI Weekly, “Fed, OCC, FDIC Press Banks on Agentic AI, Chinese Models”: https://aiweekly.co/alerts/fed-occ-fdic-press-banks-on-agentic-ai-chinese-models
- Insurance Journal, “Banking Watchdogs Ramp Up AI Scrutiny as Stability Fears Mount,” October 13, 2026: https://www.insurancejournal.com/news/national/2026/10/13/888722.htm
- ProfileNews, “AI Agents in Banks: Regulators Ask About the Kill Switch”: https://www.profilenews.com/en/ai-agents-in-banks/
- OCC Bulletin 2026-13, Interagency Revised Guidance on Model Risk Management (SR 26-2): https://aigovernance.com/policy/occ-updated-model-risk-management-guidance-2026
- Fiddler AI, “SR 26-2 Model Risk Management Guidance”: https://www.fiddler.ai/blog/sr-26-2-model-risk-management-guidance
- IT Nerd, “Regulators’ new AI questions for banks start with ‘what are you running?’,” October 14, 2026: https://itnerd.blog/2026/10/14/regulators-new-ai-questions-for-banks-start-with-what-are-you-running/
Enterprise AI Architecture
Want more enterprise AI architecture breakdowns? #
Subscribe to SuperML.