AI Governance · Field notes

Do Not Build a Second GRC Stack for Your AI Agents

By Infonaligy · Updated November 6, 2026 · 11 min read

agent controls / mapped to the frameworks you already run

Somewhere in your organization there is a slide deck titled "AI Governance Framework." It has principles on it. It probably has a maturity model. And when your SOC 2 auditor sits down next quarter and asks who approved the agent that has write access to your billing system, that deck will do nothing for you, because the auditor is not scoped to test your principles. They are scoped to test access control, change management, and audit logging. The gap between those two documents is where most enterprise AI programs are currently failing, and it is entirely self-inflicted.

The parallel program problem

Over the past eighteen months a specific pattern has become common in mid-market and enterprise IT organizations. The board asks about AI risk. Someone stands up an AI governance committee. The committee produces a policy, a risk taxonomy, and an intake form. All of that is reasonable work. The problem is what happens next: the AI governance program runs on its own track, with its own owners, its own evidence, and its own vocabulary, entirely disconnected from the compliance machinery that actually produces the reports your customers and regulators read.

When the assessment arrives, none of it converts. The auditor does not have a test procedure for "AI ethics review board attendance." They have a test procedure for logical access provisioning, and they want to know whether the non-human identity calling your CRM API was provisioned through an approved process, reviewed on a defined cadence, and deprovisioned when the project ended. If your answer lives in a governance committee's meeting minutes rather than in your identity system, you have failed a control you were probably actually satisfying in practice.

The thesis here is simple and it is not anti-governance. Do not build a second GRC stack. Take every agent-specific control requirement you have identified and translate it into the control family that already covers it in AI security and governance terms your framework already recognizes. Then be able to hand the auditor a crosswalk that shows your work.

Why 2026 forced the issue

Several things converged this year that make the parallel-program approach untenable. The EU AI Act's Article 50 transparency rules, which require AI systems that interact with people to identify themselves as AI, became applicable on August 2, 2026. The high-risk provisions scheduled to land beside them did not: the Digital Omnibus on AI deferred them to December 2, 2027 for stand-alone Annex III systems and August 2, 2028 for systems embedded in regulated products. That is a longer runway, not an exemption. For any organization with EU market exposure, disclosure is a current requirement, risk management and human oversight are dated items on the calendar, and both need to be traceable to operating controls rather than to a policy statement. We covered the disclosure side of this in more depth in our note on AI transparency and disclosure obligations.

In May 2026, CISA, the NSA, and cyber agencies from Australia, Canada, New Zealand, and the UK jointly published Careful Adoption of Agentic AI Services, and the Cloud Security Alliance followed with analyses translating it into enterprise controls. The guidance is advisory, but the technical asks in that body of work are specific: cryptographically anchored agent identities, short-lived credentials rather than static keys, encrypted agent communication, least-privilege access, and human approval for irreversible or high-impact actions. Read that list again with a control-family lens. Every single item lands inside something you already operate.

Meanwhile, the market is voting with capital. Obsidian Security raised $85 million in August 2026 at a $1.1 billion valuation specifically on demand for monitoring AI agents that touch enterprise data, and reported that nearly 70% of its clients now allow AI agents to interact with business data. On August 12, the AI Trust and Security Consortium opened applications for a founding cohort of enterprise security and technology leaders. The tooling and the peer structures are arriving. The framework guidance is arriving more slowly.

The overlay is not here, but the audit is

The most important thing an IT director should understand about the current moment is the timing mismatch at NIST. The NIST project producing SP 800-53 control overlays for these systems, renamed from COSAiS to COSIS in October 2026, published a concept paper in August 2025 and an annotated outline for predictive use cases in January 2026. As of November 2026, the overlays that cover single-agent and multi-agent deployments have not appeared even in draft, and NIST's project page gives no publication date.

That is the gap. The authoritative mapping you would ideally cite in your control narrative does not exist yet, and your next audit cycle is not going to wait for it. Some organizations have responded by pausing agent deployment pending guidance. That is not a real option for most, because the agents are already deployed, often inside SaaS products the security team never separately approved. Others have responded by declaring agents out of scope, which works exactly until the first assessor asks how a workflow automation got write access to a financial system.

The workable path is to do the mapping yourself now, document the reasoning, and treat it as a living artifact you will reconcile against the published overlay when it lands. An auditor will accept a well-reasoned, consistently applied internal mapping. An auditor will not accept silence.

Your agent controls are not new controls. They are existing controls applied to a new kind of principal.

An AI agent is a non-human identity that authenticates, holds permissions, takes actions, and produces records. Every one of those properties is already governed somewhere in SOC 2, NIST CSF, NIST SP 800-53, or ISO 27001. Agent identity is access control. Tool scope is least privilege. Action logs are audit and accountability. Prompt and model versions are configuration management. Kill switches are incident response. The work is not inventing a control catalog, it is extending your existing control descriptions to cover non-human principals and then generating evidence that proves the extension is operating. Organizations that do this give auditors something to test. Organizations that build a parallel AI program produce documents nobody tests.

The crosswalk

This is the artifact to build. The exact control identifiers will vary by which frameworks you are certified against, so treat the middle column as the family rather than a specific control number. The right column is the one that matters most in practice, because it is where teams consistently fall short: they have the control, they cannot produce evidence it was operating.

Agent controlFramework family it maps toEvidence an auditor will accept
Unique, cryptographically verified agent identity (no shared service accounts)Identity & access management; SOC 2 CC6; 800-53 IA family; ISO 27001 A.5.16Identity register showing one credential per agent, issuance and rotation records, proof that credentials are short-lived rather than static keys
Tool and API scope granted to each agentLeast privilege and authorization; SOC 2 CC6.3; 800-53 AC-6; CSF PR.AAExported permission scopes per agent, quarterly access review sign-offs naming the reviewer, denied-request logs showing enforcement actually blocks out-of-scope calls
Immutable agent action logsAudit and accountability; SOC 2 CC7.2; 800-53 AU family; ISO 27001 A.8.15Log samples showing agent identity, timestamp, tool invoked, parameters, and outcome; retention configuration; proof logs cannot be altered by the agent's own credential
Human approval thresholds for irreversible or high-impact actionsAuthorization and change control; SOC 2 CC8.1; 800-53 CM-3 and AC-3Documented threshold definitions, approval records tying a named human to a specific agent action, and evidence of what happened when an action exceeded threshold
Complete inventory of deployed agentsAsset management; CSF ID.AM; 800-53 CM-8; ISO 27001 A.5.9Current inventory with owner, business purpose, data accessed, and deployment date; reconciliation showing discovery ran against the environment, not just a survey
Model, prompt, and configuration versioningConfiguration management; SOC 2 CC8.1; 800-53 CM-2 through CM-6Version-controlled system prompts and model pins, change tickets for prompt modifications, evidence that production changes went through the same pipeline as code
Vendor-embedded agents inside SaaS productsThird-party risk management; SOC 2 CC9.2; 800-53 SR family; ISO 27001 A.5.19Vendor documentation of agent behavior and data handling, contract terms covering AI processing, and evidence the feature was reviewed before enablement
Kill switch and containment procedureIncident response; SOC 2 CC7.4; 800-53 IR-4; CSF RS.MIDocumented disable procedure with named owner and target time, plus a tabletop or live test record showing it was exercised and how long it took
Data classification and flow mapping for agent-accessible dataData classification and privacy; ISO 27001 A.5.12; 800-53 PT and SC familiesData flow diagrams showing what each agent reads and writes, classification labels on those stores, and DLP or egress controls applied to agent traffic
Escalation and fallback when the agent fails or is disabledOperational resilience and continuity; SOC 2 A1.2; 800-53 CP familyDocumented manual fallback per automated workflow, named on-call ownership, and evidence the fallback was exercised during a real or simulated outage

Two practical notes on using this. First, populate the evidence column with pointers to actual systems, not descriptions. "Okta system log, filtered to service accounts tagged agent" beats "we log agent activity." Second, expect the exercise itself to surface gaps, and treat that as the point rather than a failure. Most teams discover during mapping that they cannot tie an agent action to a verifiable identity, which is a problem worth finding before an assessor finds it. Our field notes on agent audit evidence and logging and on agent identity and access go deeper on those two rows specifically.

What genuinely does not map

Honesty matters here, because the crosswalk approach fails if you oversell it. Several agent characteristics do not fit cleanly into existing control language, and pretending otherwise will produce control descriptions that are technically satisfied and practically meaningless.

Nondeterminism breaks the testing model

Traditional control testing assumes that if you run the same procedure against the same configuration, you get the same result. Agents do not behave that way. The same prompt against the same model with the same tools can produce different action sequences. A sample of twenty-five agent transactions tells you less than a sample of twenty-five change tickets does, and no existing sampling methodology accounts for that. Nobody has fully solved this. The interim approach most mature teams use is to test the enforcement boundary rather than the behavior: you cannot prove the agent will always choose correctly, but you can prove the permission layer would have blocked it if it chose wrong.

Emergent multi-agent behavior has no control owner

When three agents each operate correctly within scope but their interaction produces an outcome none of them was authorized to produce, which control failed? Existing frameworks assume a bounded system with an identifiable owner. Delegation chains, where agent A invokes agent B which invokes a tool on behalf of the original human requester, break the chain of authorization that access control depends on. Preserving the original principal's identity through a delegation chain is a genuinely unsolved architectural problem in most enterprise deployments, and it is one reason MCP and tool-access governance deserves its own design attention rather than being folded into general API security.

Prompt injection is a novel attack class

Input validation as a control family was written for structured input. Prompt injection is an attack on the interpretation layer, where untrusted content in a document, an email, or a web page becomes instruction. There is no existing control family that cleanly covers "the data the system processed convinced it to take an unauthorized action." You can map mitigations to input validation and to boundary protection, and you should, but the mapping is uncomfortable and any honest control narrative should say so.

Proving the control was operating at the moment of action

This is the subtle one, and it is where audits will get hard over the next two years. It is not enough to show a permission scope exists today. The auditor's real question is whether the constraint was in force at the instant the agent acted. That requires correlating configuration state with action logs at timestamp granularity, which most logging architectures were never designed to support. Building it now, through deliberate agent observability and monitoring, is much cheaper than reconstructing it under assessment pressure.

Start here: a 90-day plan

The sequence below assumes you have agents in production and an audit cycle within the next year. It is deliberately ordered so that each step produces something usable even if the next one slips.

Days 1 to 30: find out what is actually running

Inventory first, always. Pull a list of every AI agent, copilot, and workflow automation with production access, and include the vendor-embedded ones inside your existing SaaS platforms, because those are the ones nobody approved. Query your identity provider for non-human identities created in the last twelve months and reconcile against what the business believes it deployed. Expect a gap. Our guide to discovering shadow AI agents covers the discovery queries in detail. Assign a named human owner to every entry. An agent without an owner is an unassigned risk, and it will be the first thing an assessor finds.

Days 31 to 60: build the crosswalk and find the evidence gaps

Take the table above, adapt it to the frameworks you are certified against, and for each row write down the specific system, query, or report that would produce the evidence. Do not write what you intend to build. Write what exists today, and mark the empty cells. Those empty cells are your remediation backlog, priced and sequenced. Expect the empty cells to cluster in the same three places: agent action logging that lacks identity attribution, permission scopes that were granted once and never reviewed, and complete absence of any tested disable procedure.

Days 61 to 90: close the highest-risk gaps and test one end to end

Fix identity attribution in logs first, because it is a prerequisite for almost every other piece of evidence. Then run one real containment test: pick a production agent, disable it, time how long it takes, confirm the fallback process works, and document all of it. That single artifact does more for an assessment than a quarter of committee meetings. Update your existing policies to explicitly cover non-human principals rather than writing a new AI policy, and align the language with your agent security policy and your broader governance checklist. If you want an outside read on where you stand before committing budget, an assessment will surface the gaps faster than an internal review will, because internal reviews tend to grade the systems people built.

What this looks like when it is working

A mature program does not look impressive from the outside. There is no AI governance portal. What there is: agents appear in the same asset inventory as servers, agent credentials appear in the same access review as employee accounts, prompt changes move through the same change advisory process as code, and agent actions appear in the same SIEM as everything else. When the auditor asks about AI, the security lead pulls up a two-page crosswalk, points at the control families, and produces the same evidence artifacts they produce for everything else.

That outcome is mostly an engineering problem rather than a policy problem. It requires identity infrastructure that treats agents as first-class principals, logging that attributes actions correctly, and enforcement at the tool boundary rather than in the prompt. Teams building agents in-house have an advantage here, because these properties can be designed in from the start rather than retrofitted. If you are scoping new custom AI agents or extending existing automation, the control requirements belong in the architecture review, not in a compliance review six months later.

The practical takeaway

Stop treating AI governance as a new discipline and start treating agents as a new type of principal inside the disciplines you already run. Build the crosswalk. Populate the evidence column with real system pointers. Be honest in your control narratives about the four or five things that genuinely do not map yet, because a documented known gap reads far better in an assessment than an overstated claim that falls apart under testing. And do the mapping now rather than waiting for the NIST overlays, because the audits are already happening and a defensible internal mapping is what carries you until the authoritative one publishes. Infonaligy works with companies across Dallas–Fort Worth and remotely nationwide on exactly this problem: consulting engagements that inventory what is running, map it into the frameworks you already maintain, and close the evidence gaps before someone else finds them. You can see where we operate on our locations page.

Infonaligy is an AI-native managed intelligence provider based in Dallas–Fort Worth, working with security, compliance, and IT teams across Texas, Oklahoma, and remotely nationwide.

Audit-ready agent governance

Get your agent controls mapped before the auditor asks.

Infonaligy builds the crosswalk that connects your deployed agents to the SOC 2, NIST, and ISO 27001 control families you already maintain, then closes the evidence gaps it exposes. We inventory what is actually running, fix the identity and logging problems, and leave you with artifacts your assessor accepts.

Agent control mapping · DFW · remote nationwide · audit-ready · 800-985-1365