Skip to content
RiskTemplates · The Daily Brief Thursday, November 19, 2026
Wire FINRA's Spartan Capital Case: The Private-Placement Diligence File That Wasn't There NOV 17

Feature AI Risk

Your Agent Is Logged In as a Service Account Nobody Owns

SR 26-2 pushed agentic AI out of model risk scope. AI agent access control now needs an entitlement register, JIT credentials, and a named owner.

By Rebecca Leung · November 18, 2026 ·
Table of Contents
  • SR 26-2 spends a paragraph pushing generative and agentic AI out of model risk scope. Most teams read that as “not us yet.” Read the next sentence instead: non-compliance with the guidance won’t draw supervisory criticism, but violations of law and unsafe practices still will.
  • The gap the carve-out opens is not conceptual. It’s operational, and it lives in your IAM system, where an agent is usually running as a service account that a contractor created and nobody has looked at since.
  • Canada’s banking regulator published the control list in July 2026 that US guidance hasn’t: unique non-human identities, least privilege, just-in-time credentials, tool allow-lists, approval checkpoints, and periodic recertification.
  • An AI use-case inventory won’t get you there. You need the entitlement columns — scope, credential lifetime, approver, certifier, retirement trigger — and an owner who is a person.

Here is the sentence, from page 3 of the SR 26-2 attachment: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.”

That’s it. One carve-out, no roadmap, issued April 17, 2026 by the Federal Reserve, OCC, and FDIC together with the SR 26-2 transmittal that retired SR 11-7 and SR 21-8.

The predictable response inside banks was relief. Agents are out of scope, so agents are somebody else’s problem until the agencies write the agentic guidance. That reading is half right and operationally useless, because the thing that makes an agent dangerous isn’t statistical — it’s an entitlement.

What the carve-out actually does

SR 26-2 is explicit that it binds nobody. The attachment says the guidance “does not set forth enforceable standards or prescriptive requirements; accordingly, non-compliance with this guidance will not result in supervisory criticism against a banking organization.”

Then the next clause: supervisory action may still follow from “any violations of law or unsafe or unsound practices stemming from insufficient management of model risk.”

So the carve-out doesn’t move agents into a regulatory void. It moves them into the older, blunter world of unsafe practice — where the question isn’t whether you followed a framework, it’s whether a reasonable firm would have let this thing have that access. It also narrows its own model definition to exclude “deterministic rule-based processes and software with no statistical, economic, or financial theory underpinning design or use,” which means the argument that an agent is “just automation” gets you excluded from the guidance and included in the risk.

Meanwhile the guidance applies most directly above $30 billion in assets and still touches smaller institutions with significant model exposure. If you’re under the threshold, you don’t get a pass — you get the same carve-out with less infrastructure to absorb it.

The failure mode is inheritance, not provision

Ask how a new agent gets access and the honest answer at most institutions is: someone requests it in the same workflow they’d use for a human, a service account gets created or reused, and a scope gets attached.

Three specific things go wrong.

Shared credentials. The agent runs as an existing service account because standing one up takes four days and the sprint ends Friday. Now the agent and a batch job and a legacy integration share one identity. When an alert fires, you cannot attribute the action. NIST’s NCCoE named shared credentials and long-lived access as exactly the risk in its February 5, 2026 concept paper on agent identity and authorization.

Scope by convenience. Provisioning grants what the integration test needed, which is usually read-write to the whole object rather than the fields the use case touches. Nobody narrows it afterward because narrowing requires a regression test.

Tool chaining. A single scoped permission set is fine until the agent can call a tool that calls another system. Canada’s banking regulator, OSFI, listed tool chaining alongside over-privileged access and weak traceability in its July 2026 technology risk bulletin. The bulletin’s framing is worth borrowing verbatim in your own risk assessment: agentic AI “amplifies Gen AI risks by acting on unvalidated outputs.”

That last point is the compound failure. A hallucination that lands in a chat window is an embarrassment. A hallucination that lands in a write call is a data integrity event with a timestamp attributable to a credential you can’t name the owner of.

The register an examiner will ask for

An AI use-case inventory tracks intent. An agent entitlement register tracks authority. You need both, and they should reconcile — the two-register reconciliation problem already exists in your model risk stack, and it gets worse with agents because the join key is a credential rather than a model ID.

Minimum viable schema, one row per agent:

FieldWhy it existsExample entry
agent_name / purposeTies the identity to a decisionap-recon-agent — matches remittance exceptions to invoice lines
owning_humanThe person, not the teamNamed ops manager, with a named backup
accountable_execEscalation and acceptance authorityDirector, Payment Operations
identity_typeDedicated NHI vs. inherited service accountDedicated NHI, created 2026-09-14
systems_and_scopeExact entitlements with approval dateap_ledger read; remit_queue write; approved 2026-09-14
credential_modelLifetime and issuanceJIT, 15-minute tokens, vault-issued
tool_allowlistChained-call boundarylookup_invoice, post_match, no approve_payment
checkpointWhere a human must confirm>$25k variance routes to ops review
log_destinationTraceabilitySIEM index, 13-month retention
last_recertDrift check2026-11-03, business certifier
retire_triggerAnti-zombie control90 days idle, or pilot end date

The identity_type column is the one that produces findings, because populating it forces the question of whether the agent borrows a credential. Any agent sitting on an inherited service account is a remediation item on day one. Not a project. An item.

Ownership is where this gets ugly

Agent governance sprawls across three teams with incompatible incentives. IAM owns identity lifecycle and will provision whatever the ticket says. AI governance owns the use case and cares whether the model is fit for purpose. The business owns the outcome and cares that reconciliation stops taking four hours.

Nobody owns the intersection, and the intersection is where every finding lives.

A workable split, tested against what an examiner can actually inspect: IAM owns issuance, lifetime, and vaulting. AI governance owns the register and runs recertification. The business unit certifies each entitlement as still necessary. Internal audit samples the register against the IAM export quarterly and reports exceptions to the AI risk committee. The artifact that proves the model works is the reconciliation between register rows and active directory entries — if an agent is live but unregistered, the control failed and everyone can see it.

Recertification cadence should track privilege, not calendar convention. Standing high-privilege agents monthly. Everything with write access to a system of record quarterly. Read-only agents quarterly at minimum until you have a year of clean data. A workable starting point is to take your highest-scope agent and run the first cycle this month so the cadence has a baseline to calibrate against.

The real drift risk isn’t the initial grant. It’s the agent that outlives its project because the decommission step was never wired into the offboarding checklist — the same pattern that bulletin 2026-13 put pressure on for traditional models.

The kill switch needs a rehearsal date

Every agentic deployment needs a documented stop: who can revoke, how fast, what state is left behind, and how in-flight work is reconciled. Rehearse it once per quarter and record the time to revocation.

The reason is not theoretical. If an agent posts bad matches for six hours, the first question is when it stopped and the second is what it touched. A revocation that takes a change ticket and two days means the second question has a two-day answer.

Include the rollback path for the data, not just the identity — the last committed batch, the reversal script, the counterparties to notify. Identity revocation without a data reversal stops the bleeding at the wrong end.

So what this changes

Reading the SR 26-2 carve-out as a deferral buys you a year of drift. During that year, the agent population grows, the service accounts multiply, and the register stays empty. Then the next exam asks for it and the answer is a spreadsheet built the week before.

The build is small and unglamorous: an entitlement register with a human in every owner field, JIT credentials, a tool allow-list, a quarterly rehearsal, and an audit sample. Do the identity_type census first — count how many agents ride an inherited service account. That number is your finding, and it’s better to find it yourself.

The AI Risk Assessment Template includes the agentic addendum with the register columns above, the tool-chaining worksheet, and a revocation rehearsal log — it’s here. Start with the census, not the policy.

FAQ

Does SR 26-2 cover agentic AI?

No. The April 17, 2026 guidance states that “generative AI and agentic AI models are novel and rapidly evolving” and “are not within the scope of this guidance.” It covers traditional quantitative models and non-generative, non-agentic AI. The carve-out is a scope statement, not an exemption.

Is SR 26-2 enforceable?

Not on its own — it “does not set forth enforceable standards or prescriptive requirements.” But the same paragraph preserves supervisory action for violations of law or unsafe or unsound practices stemming from insufficient model risk management. An agent that mis-allocates is an unsafe practice whether or not the guidance reaches it.

What belongs in an AI agent register?

Purpose, owning human, accountable executive, identity type, every system scope with approval date, credential lifetime, tool allow-list, human checkpoint, log destination, last recertification, and a retirement trigger. The identity-type column matters most because inherited service accounts break attribution.

How often should agent entitlements be recertified?

Standing high-privilege agents monthly; anything with write access to a system of record quarterly with a named business certifier. The risk is drift, not initial provisioning — credentials outlive projects when decommissioning never reaches the offboarding checklist.

What does OSFI recommend for agent access?

Its July 2026 bulletin lists over-privileged access, shared credentials, tool chaining, weak traceability, and automated execution as core risks, and recommends unique non-human identities, least privilege, scoped permissions, just-in-time access with short-lived credentials, tool allow-lists, approval checkpoints, activity logging, and periodic recertification.

◆ Need the working template?

Start with the source guide.

These answer-first guides summarize the required fields, evidence, and implementation steps behind the templates practitioners search for.

◆ Immaterial Findings · Weekly

Sharp risk & compliance insights. No fluff.

◆ FAQ

Frequently asked questions.

Does SR 26-2 cover agentic AI?
No. The April 17, 2026 guidance says plainly that 'generative AI and agentic AI models are novel and rapidly evolving' and 'are not within the scope of this guidance.' It covers traditional statistical and quantitative models plus non-generative, non-agentic AI. The carve-out is a scope statement, not an exemption — supervisory action can still follow from violations of law or unsafe practices.
Is SR 26-2 enforceable?
Not on its own. The attachment states the guidance 'does not set forth enforceable standards or prescriptive requirements' and that non-compliance 'will not result in supervisory criticism.' But the same paragraph preserves supervisory action for violations of law or unsafe or unsound practices stemming from insufficient model risk management. An agent that misprices or mis-allocates is an unsafe practice whether or not the guidance reaches it.
What should an AI agent register contain?
Agent name and business purpose, owning human and accountable executive, the identity type (dedicated non-human versus inherited service account), every system and data scope granted with the approval date, credential lifetime and access model, tool allow-list, approval checkpoint for high-impact actions, logging destination, last access recertification, and the retirement date or review trigger. An AI use-case inventory usually has none of the entitlement columns.
How often should agent entitlements be recertified?
Set it from privilege, not from convention. Standing high-privilege agents monthly, all others quarterly, and any agent with write access to a system of record at least quarterly with a named business certifier. The failure mode is drift, not initial provisioning — agent credentials routinely outlive the project that created them because nobody scheduled the review.
What does OSFI actually recommend for agent access?
Its July 2026 technology risk bulletin on generative and agentic AI lists over-privileged access, shared or inherited credentials, tool chaining, weak traceability, and automated execution as the core risks, and recommends unique non-human identities, least privilege, scoped permissions, just-in-time access with short-lived credentials, tool allow-lists, approval checkpoints on high-impact actions, activity logging, and periodic access recertification.
Rebecca Leung

Author

Rebecca Leung

Rebecca Leung has 8+ years of risk and compliance experience across first and second line roles at commercial banks, asset managers, and fintechs. Former management consultant advising financial institutions on risk strategy. Founder of RiskTemplates.

◆ Related framework

AI Risk Assessment Template & Guide

Comprehensive AI model governance and risk assessment templates for financial services teams.

Immaterial Findings · Newsletter

The brief, in your inbox.

Enforcement of the week, a framework breakdown, and the prompts that are actually worth running. Delivered to your inbox. Free.