AI Agent Security Review Checklist: Ten Questions Before Launch
Ten questions an IT or operations lead should answer before an AI agent goes live with company data and tools, plus a practical pre-launch plan.

A pre-launch security review for an AI agent is a short written check that confirms what the agent can read, what it can change, which steps need a person to approve, and how you stop it. An IT or operations lead should be able to answer the ten questions below before the agent receives company data or live tool access. If the honest answer to any of them is that nobody has decided yet, the agent is not ready for production.
Key takeaways
- Treat the agent as a new user with its own identity, named owner, credentials and audit trail, not as a feature of a chat window.
- Write down the data it may read, the tools it may call and the actions that need a person to approve before they run.
- Test the stop path and the staff override before launch, not after the first unexpected action.
- Keep records of decisions and blocked actions, and keep personal data out of those records unless review requires it.
- Define one observable success criterion so you can tell whether the agent behaves as designed next month, not only on launch day.
Why an agent review is not a chatbot review
A chatbot produces text that a person reads before anything happens. An agent reads records, calls tools, writes to systems and sometimes carries memory from one task to the next. That changes the risk. A wrong sentence can be corrected in the next message, but a wrong update to a customer record, a message sent to the wrong contact or a file shared with the wrong group may already have consequences by the time someone notices.
The main risk categories are now documented in public frameworks. The NIST AI Risk Management Framework gives organizations a structure for mapping, measuring and managing AI risk, and the OWASP Top 10 for LLM Applications lists the failure patterns that appear most often in language model applications, including prompt injection and excessive permissions. You do not need to adopt either document in full before launch. You do need to answer the questions they point to, in writing, for the specific agent you plan to run.
Questions 1 to 3: Identity, purpose and boundaries
The first three questions establish what the agent is and what it is for. Most later problems trace back to a vague answer here.
- Question 1: Does the agent have its own account and credentials? It should not operate through a shared login or through the personal token of an employee, because then its actions cannot be separated from that person's actions in the logs.
- Question 2: What single job is it hired for, and what is explicitly out of scope? Write the job as one sentence and list at least three things the agent must not do, such as issuing refunds, changing prices or answering legal questions.
- Question 3: Who is accountable for its behavior? Name one business owner who decides scope and one technical owner who maintains access. If the names are missing, the review is not finished.
A useful test is to ask whether a new staff member could read the written job description and know what to do and not do. If the agent's description would confuse that staff member, it will confuse the agent too.
Questions 4 to 6: Data, memory and tools
These questions set the surface the agent can reach. Start with a list, not an intention.
- Question 4: What data can it read, and what data must it never see? Separate customer records, employee records, financial documents and contracts. Apply data minimization: give the agent only the fields a task requires.
- Question 5: What memory does it keep, where is that memory stored, and how is it deleted on request? Memory that holds personal details from a conversation can outlive the conversation unless someone defined retention and deletion rules.
- Question 6: Which tools can it call, with which permissions, and which of those calls write data? Each tool should use scoped credentials that grant only the operations the job needs. The OWASP Authorization Cheat Sheet is a practical reference for checking that access decisions happen on the server side and are not left to the model's judgment.
For the tool list, mark every operation as read or write, then ask whether a write can be undone. Read-only access is easier to justify at launch. Each write that remains should have a named reason. Schedule a permission review for the period after launch, since tools and staff roles change. The quarterly access review process shows how to compare granted permissions against the job that is actually running.
Questions 7 and 8: Approvals and reversibility
Question 7 asks which actions require a human decision before they run. Build a short table in your head with three columns: actions the agent may take alone, actions it may propose but not execute, and actions it must never attempt. Sending a reminder to a customer may belong in the first column. Changing a contract amount, deleting a record or sharing a file externally usually belongs in the second. The human approval gates guide explains where people should sign and how to keep those gates from becoming rubber stamps.
Question 8 asks whether a risky action can be reversed and what happens to queued work when you pause the agent. A pause should stop new actions without deleting logs, dropping pending tasks silently or leaving half-finished writes behind. Test this before launch, with a real pending task. The kill switch guide describes how to halt actions while keeping the audit trail and the queue intact.
Questions 9 and 10: Logs, tests and incident response
Question 9 asks what the agent records, who can read those records and how personal data is kept out of them. Logs are necessary for review, but raw logs can copy sensitive content into places that were never meant to hold it. Decide which fields are recorded, which are masked and how long the records are kept. The PII redaction guide covers how to keep audit detail without creating a second copy of the data.
Question 10 asks how you will know the agent is failing and who responds. Define at least one success criterion that can be observed, such as the share of tasks completed without staff correction, measured against a fixed set of test cases. Then define the failure signals: unexpected write attempts, repeated blocked actions, complaints from customers or unusual tool usage. Name the person who receives each signal and the steps they take. Testing with realistic prompts, including attempts to steer the agent into unauthorized actions, belongs here. The OWASP prompt injection prevention cheat sheet lists defensive patterns that are worth checking against your design.
Illustrative scenario and a launch checklist
Illustrative: a regional equipment distributor wants an agent that answers order status questions and opens return requests in its order system. In the review, the team discovers that the first draft of the integration uses a credential that can also edit prices. The scope is cut to read access plus creating return requests. Refund approval stays with a staff member, and the agent's records of blocked actions are reviewed each week during the first month. Nothing about the model changed. The boundaries did.
Before launch, confirm each item below in writing:
- The agent has its own account, a business owner and a technical owner.
- The job statement lists what the agent must not do.
- Every data source is listed, with the fields the agent may read.
- Memory retention and deletion rules are documented.
- Every tool is marked read or write, with scoped credentials.
- Actions needing approval are listed, and the approver is named.
- The stop path was tested with a pending task.
- Logs are defined, with personal data masked where required.
- Success criteria and failure signals have named recipients.
How Be Human runs this review
Be Human designs, implements and operates personalized AI infrastructure in which each agent keeps its identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion, so these questions are answered during design rather than discovered after launch. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission review and incident response, which keeps the written answers accurate as systems change.
If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Do small companies need a formal AI agent security review? Yes, although it can be short. A company with a few staff and one agent still needs written answers about identity, data, tools, approvals and the stop path. The review can fit on a few pages, but the decisions behind it should be explicit and owned by a named person.
Can we reuse our existing IT security checklist for an AI agent? Partly. Your existing checklist likely covers accounts, permissions, logging and incident response, which all apply. It usually does not cover memory retention, prompt injection, approval gates for model-proposed actions or the behavior of a system that decides which tool to call next, so those items need to be added.
Who should sign off before an agent goes live? The business owner who is accountable for the outcome should sign off on the scope and approval rules, and the technical owner should sign off on credentials, data access and logging. Where personal data is involved, the person responsible for privacy compliance should also review the data list and retention rules before launch.
How often should the review be repeated? Repeat the full review before any significant change, such as a new tool, a new data source or a wider scope. Between those changes, run a lighter permission and log review on a recurring schedule, because staff roles, vendor integrations and business rules tend to shift slowly and the gap between documented and actual access grows without anyone noticing.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


