The October 2026 issue of the Journal of Accountancy carries an article headed “Cyber secure: New checklist helps CPAs manage risks”. As described in coverage of the piece, it sets out a cybersecurity baseline for accounting and finance professionals across 12 core control areas, and it brings AI governance directly into the wider control environment rather than treating it as a separate topic.
Our Crest.Digital News Bite explores the growing convergence of AI governance, cyber risk and internal audit. Watch it below, then read on for a practical way to apply the idea.
One caveat: the full article sits inside the Journal’s digital edition, so the list of measures in this piece reflects how the checklist has been described to us, not a reproduction of its text. Refer to the Journal for the authoritative wording.
Crest.Digital’s Agentic GRC builds configurable AI agents for control testing, evidence collection and remediation tracking, with human-in-the-loop review and an audit trail on every action.
Explore Agentic GRCWhat the Checklist Signals
Checklists are useful precisely because they are plain. This one is aimed at the people who run finance and reporting processes, and the message is that AI governance is no longer a specialist side project. It belongs in the same baseline as access control, logging and recovery planning. That mirrors how frameworks such as the NIST Cybersecurity Framework and the COSO internal control framework already treat technology risk: as part of the control environment, tested in the ordinary course.
Why AI Agents Are a New Control Surface
A traditional user logs in, does a task and logs out. An AI agent may hold standing credentials, call tools and APIs, read ledgers and trigger actions at machine speed. That makes it behave like a privileged user, but one that is often missing from the access register. Our piece on AI agent identity registers and non-human privileged users covers the gap in detail.
So the question is practical, not philosophical. If an annual assessment tests user access, change management and backups but not the agents that now operate inside the same systems, it is assessing yesterday’s control surface. The same logic applies to fraud: agents that can initiate or approve payments need the same dual-control discipline as people, which we discuss in AI agents and human approval in financial reporting.
The Measures Highlighted, Grouped for Testing
Twelve Measures Internal Audit Can Test
- Human oversight of AI: Named owners and approval points for consequential agent actions.
- Role-based access controls: Agent and user permissions tied to defined roles.
- Least-privilege access: Each agent can reach only what its task requires.
- Activity logging: Agent and user actions recorded in a form a reviewer can replay.
- Multifactor authentication: Enforced for users and administrators of AI tools.
- Restricted tools and APIs for AI agents: Allow-lists rather than open access.
- Monitoring of AI activity: Alerts for unusual agent behaviour, with follow-up.
- Secure operating environments: Hardened, segregated environments for agents and data.
- Resilience and recovery planning: Tested backups and a plan for AI-dependent processes.
- Fraud prevention controls: Controls designed for AI-enabled impersonation and payment fraud.
- Dual authorization for major transactions: Two approvers, whether the initiator is a person or an agent.
- Evidence and accountability: Records that show who did what, when and under which approval.
Notice how many of these are classic IT general controls applied to a new kind of actor. That is encouraging: most teams do not need a new methodology, only an extended scope. The harder part is evidence, since agents change faster than annual cycles, a gap discussed in why compliance does not equal security.
A Practical Assessment Workflow
The eight steps below turn each measure into a repeatable assessment. They are written for internal audit but work equally well for second-line teams and for finance leaders running self-assessments.
Control Requirement
State each control as a testable requirement, for example that every AI agent has a named owner and least-privilege access. A vague principle cannot be tested.
Evidence
Define in advance what proves the control operates: access reviews, logs, configuration exports and approval records. See our guide on agentic internal audit and evidence validation.
Test
Perform the test, whether inspection, re-performance or automated check. For agents, include their tool access, API scope and logging, as set out in AI agent identity and access governance.
Finding
Record exceptions with the control, the evidence gap, the cause and an accountable owner, not just a pass or fail.
Risk Rating
Rate each finding by consequence and likelihood, considering what the agent or user can reach, so remediation effort follows risk.
Action
Assign a remediation action, owner and due date, with escalation if the date slips.
Closure
Close a finding only with evidence that the fix works, kept with the record.
Retest
Test again after remediation and on a recurring cycle. Continuous checks catch drift between annual assessments, a theme in moving from sampling to continuous assurance.
The test is traceability. For any measure on the list you should be able to show the requirement, the evidence, the test, any finding, its rating, the action taken, the closure proof and the retest result. Where one is missing, assurance is incomplete.
Where Agentic AI Helps, and Where People Stay Accountable
Testing twelve areas across many systems and agents is repetitive work, which is where agents help: gathering access exports, comparing permissions against approved roles, running scheduled checks, drafting findings and chasing owners. What they must not do is certify their own controls. Separation of duties still applies, and the logging expectation applies to the testing agents too, as we argue in why AI agents need audit trails and in our note on AI governance moving toward independent assurance.
Crest.Digital’s Voice AI for GRC follows up with control owners for evidence and remediation updates, and the TPRM platform extends the same assurance discipline to the vendors behind your AI.
Five Moves for the Next Control Cycle
A Short Playbook for Audit, Finance and Risk Leaders
- Add AI to scope: Include agents and AI tools in the annual control universe, tiered by what they can access and change.
- Register the agents: Keep an inventory with owners, permissions, tools and approval points.
- Reuse existing tests: Extend access, logging, MFA and recovery tests to agents before inventing new ones.
- Test between cycles: Schedule lighter, automated checks so drift does not wait for the annual review.
- Report coverage: Show the audit committee how many agents are covered, what was found and what has closed.
A final note on proportion. A checklist is a starting baseline, not a guarantee, and not every AI tool needs full testing. Tier by consequence and document your reasoning. Our business impact page shows how organizations see results from evidence-led control programmes.
Crest Perspective
Cyber risk in vendor relationships is not uniformly distributed. The vendors that represent the highest exposure are rarely those with the longest questionnaire responses — they are the ones with privileged access to production systems, those processing unstructured data at scale, and those operating in jurisdictions with limited breach notification requirements. Risk tiering that reflects actual access and data flows is more predictive than vendor size or certification status alone.
Frequently Asked Questions
The Journal of Accountancy's October 2026 issue includes an article titled Cyber secure: New checklist helps CPAs manage risks, describing a cybersecurity baseline for accounting and finance professionals across 12 core control areas, including AI governance, cyber resilience and security controls. Measures highlighted in coverage of the checklist include human oversight of AI, role-based and least-privilege access, activity logging, multifactor authentication, restricted tools and APIs for AI agents, monitoring of AI activity, secure operating environments, resilience and recovery planning, fraud prevention, dual authorization for major transactions, and evidence and accountability. Refer to the Journal itself for the authoritative wording.
Increasingly, yes. As AI agents gain access to systems, data and workflows, they become a control surface alongside traditional IT general controls, cybersecurity and operational controls. An annual assessment that tests access, change management and backups but ignores the agents that can now read ledgers, send payments or modify records leaves a growing blind spot. Depth should be risk-based: agents that touch money, customers or regulated data need testing; a drafting assistant with no system access needs far less.
Treat each agent like a privileged user. Test whether it has a named owner, whether its access is limited to what its task needs, whether its tools and APIs are restricted, whether its activity is logged and monitored, whether humans approve high-risk actions, and whether major transactions still require dual authorization. For each control, define the requirement, collect evidence, perform a test, record findings, rate the risk, assign action, confirm closure and retest. Independent testing matters: the team that runs the agent should not be the only team assessing it.
Evidence should show that the control operated, not merely that a policy exists. Examples include access review records showing agent permissions were checked, logs showing agent actions and approvals, configuration exports showing multifactor authentication is enforced, tool and API allow-lists, monitoring alerts with follow-up, recovery test results and approval records for large transactions. Evidence is strongest when captured as the control operates, linked to the requirement, and retained so a reviewer can reproduce the conclusion.
Agents can handle the repetitive groundwork: collecting evidence, running scheduled tests, comparing access against approved roles, drafting findings and chasing remediation. Independence is protected through separation of duties, so agents that operate a system do not certify its controls, and through logging every agent action so a reviewer can reproduce results. Crest.Digital's Agentic GRC follows this human-in-the-loop model, with accountable people rating risk and signing off.