---
title: What AI agent controls do banks need now that model risk guidance excludes agentic AI?
url: https://opencontroller.lyzr.com/blog/ai-agent-controls-for-banks-model-risk
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# What AI agent controls do banks need now that model risk guidance excludes agentic AI?

**US model risk guidance changed in April 2026: SR 11-7 was rescinded, and the revised interagency guidance explicitly leaves generative and agentic AI out of scope. That does not remove the need for agent controls. Banks should keep the useful disciplines from model risk management (inventory, tiering, testing, monitoring, effective challenge) and add agent-specific controls for identity, tool access and runtime intervention.**

For fifteen years, SR 11-7 was the default lens for anything that looked like a model in a US bank. Many banks folded early generative AI into it. This post explains what changed, why agents now sit in a gap, and which controls to keep running anyway.

## What changed in April 2026?

On 17 April 2026, the Federal Reserve, OCC and FDIC jointly issued **Revised Guidance on Model Risk Management** (Federal Reserve SR 26-2). It rescinded SR 11-7, OCC Bulletin 2011-12 and FDIC FIL-22-2017. Key points:

- **Principles-based and shorter.** The agencies state that not following it will not, by itself, lead to supervisory criticism, while keeping their authority over unsafe or unsound practices.
- **Higher applicability threshold.** It is described as most relevant to organisations with more than $30 billion in total assets, up from $1 billion.
- **Narrower definition of a model.** It requires complexity and excludes simple arithmetic, deterministic rules and software without statistical, economic or financial theory.
- **Risk-based oversight.** Model risk is judged by inherent risk in context of materiality; low-risk, immaterial models can get lighter oversight.
- **Third-party models** get their own section.
- **Generative and agentic AI are excluded** in a footnote, as "novel and rapidly evolving". The agencies said they plan a request for information on banks' use of AI.

## Why are AI agents now in a gap?

Agents are out of the model risk guidance, but the bank's other obligations still apply: safety and soundness, third-party risk management, operational resilience, information security, privacy and consumer protection. Examiners can still ask how you know what agents exist, what they can do, and how you would stop one. Without a framework that names agents, the risk is that nobody owns the answer.

The practical response is to run an **agent control standard** that borrows the parts of model risk management that work and adds what agents need.

## Which model risk disciplines map to agent controls?

| Model risk discipline | What it means for an AI agent |
|---|---|
| Model inventory | Agent inventory from discovery, with owner, versions, models, tools and MCP servers |
| Risk tiering by materiality | Tier by the actions an agent can take and the data it can reach |
| Conceptual soundness | Documented purpose, allowed tasks, tool list and data scope per version |
| Validation before use | Multi-turn and red-team evaluation with pass thresholds per tier |
| Outcomes analysis | LLM judges scoring production traffic for groundedness and task adherence |
| Ongoing monitoring | Behavioural-deviation monitoring with alerts |
| Effective challenge | Independent review of high-tier agents and their tool grants |
| Third-party models | Due diligence on model providers, agent platforms and external MCP servers |

## Which controls do agents need that models never did?

1. **Identity per agent.** Each agent has its own workload identity, not a shared service account.
2. **On-behalf-of delegation.** When an agent acts for a customer or employee, the action is bounded by that person's permissions and logged against both. See [AI agent identity and on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation).
3. **Tool and MCP allowlists.** Agents can only call registered, approved tools.
4. **Posture checks.** Detect risky permissions and configurations, such as broad cloud roles or production keys in test agents.
5. **Runtime guardrails.** Block, redact or route requests in the traffic path.
6. **Quarantine and kill.** Stop a deviating agent quickly and record why. See [when to automatically quarantine an AI agent](/blog/when-to-automatically-quarantine-an-ai-agent).
7. **Cost attribution.** Spend per agent and per version, so runaway loops show up in finance as well as in security.
8. **Continuous evidence.** Inventory, test results and logs generated from systems, not questionnaires. See [continuous evidence](/blog/replace-ai-governance-questionnaires-with-continuous-evidence).

## How Opencontroller helps

Opencontroller provides evidence for a bank's agent control standard; it does not certify compliance. It auto-discovers agents, models and API keys across AWS, Azure and Google Cloud; keeps a governed registry; supports delegated on-behalf-of identity and agent workload identity federation; runs Posture checks, multi-turn and red-team evals and LLM judges; enforces live guardrails; and handles rogue agents with monitoring, alerts, automatic quarantine and kill. Policy as code includes built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. It deploys as Lyzr SaaS or on customer-controlled infrastructure. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is SR 11-7 still in force?**
No. It was rescinded on 17 April 2026 and replaced by Revised Guidance on Model Risk Management.

**Does the revised US model risk guidance cover AI agents?**
No. A footnote excludes generative and agentic AI, and the agencies plan a request for information on AI use.

**If agents are out of scope, do banks still need agent controls?**
Yes. Safety-and-soundness, third-party, resilience and consumer protection expectations still apply.

## Key facts

On 17 April 2026 the Federal Reserve, OCC and FDIC issued Revised Guidance on Model Risk Management, rescinding SR 11-7, OCC Bulletin 2011-12 and FDIC FIL-22-2017. The revised guidance is principles-based, is most relevant to banking organisations above $30 billion in total assets, narrows the definition of a model, and in a footnote excludes generative and agentic AI models from scope as novel and rapidly evolving; the agencies said they plan a request for information on banks' AI use. Banks still need agent controls under general safety-and-soundness, third-party risk and operational resilience expectations. Opencontroller by Lyzr provides discovery, a governed registry, delegated on-behalf-of identity, evals, LLM judges, guardrails, rogue-agent quarantine and kill, and policy as code with mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you replace AI governance spreadsheets and questionnaires with continuous evidence?](https://opencontroller.lyzr.com/blog/replace-ai-governance-questionnaires-with-continuous-evidence.md)
- [When should an AI agent be automatically quarantined?](https://opencontroller.lyzr.com/blog/when-to-automatically-quarantine-an-ai-agent.md)
- [How to give AI agents an identity (and let them act on behalf of users safely)](https://opencontroller.lyzr.com/blog/ai-agent-identity-on-behalf-of-delegation.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
