- CISO
- CIO
Would You Let This AI Agent Go Into Production?
THE SHORT ANSWER
Before an AI agent goes into production, ask for two pieces of evidence: its actual permissions in every system it connects to, and the result of a stop test showing what happens when an administrator pauses it. If either is missing, the agent isn't ready, however good the demo looked.
On this page
A sales operations team builds an agent that prepares a weekly pipeline report from the CRM. In the demo it works well.
Before sign-off, security asks to see its permissions. The agent runs under a team member's account, so it can edit opportunities and send email outside the company, not just read the report fields. Nobody has tried switching it off.
Whose access does the agent use?
Name a business owner and the administrator responsible for the agent. Then list every system it connects to and the identity it uses in each. It might act as the signed-in user, use a shared integration account, or have an identity of its own.
Each choice carries risk. OWASP's Non-Human Identities Top 10 (2025) lists "Overprivileged NHI", "NHI Reuse" and "Human Use of NHI" among the main problems. An agent running under a person's account inherits everything that person can do. A dedicated identity with only the access the job needs is easier to limit, monitor and remove.
How much access should it get?
Only what the job requires. OWASP's guidance on excessive agency says to "Limit the permissions that LLM extensions are granted to other systems to the minimum necessary" and to "require a human to approve high-impact actions before they are taken."
For the sales report, that means read-only access to the CRM fields the report uses. Leave out record editing and external email unless the workflow needs them. If the report must go outside the company, a person approves the recipients and content first. Apply the limit in the connected system itself, and check that it holds.
How do you know you can stop it?
Run a stop test in a test environment or a controlled pilot:
- Have the administrator pause the agent.
- Check what happens to work that was queued or already running.
- Confirm the agent can no longer use any of its connections.
- Check that the activity logs show what it did before it stopped.
Platforms handle this differently. If pausing leaves a connection active, write down who can revoke that access and what else depends on the same connection.
How do the gates map to OWASP's agentic risks?
OWASP published its Top 10 for Agentic Applications on December 9, 2025. Five of its risks map directly to the two gates:
| OWASP risk | What it looks like in practice | Which gate catches it |
|---|---|---|
| ASI01 Agent Goal Hijack | Instructions hidden in a document or email redirect what the agent tries to do | Gate 1: tight permissions limit the damage, and human approval blocks high-impact actions |
| ASI02 Tool Misuse | The agent uses a connected tool for more than it was meant to | Gate 1: permissions reviewed in each system |
| ASI03 Identity & Privilege Abuse | The agent runs with a person's or a shared admin account | Gate 1: a dedicated identity with least privilege |
| ASI04 Agentic Supply Chain Vulnerabilities | A third-party tool, plugin or model the agent relies on changes or is compromised | Gate 1: an inventory of connections, and a review when they change |
| ASI10 Rogue Agents | The agent keeps acting when it shouldn't, and nobody can stop it | Gate 2: a tested stop |
Where should you start?
Pick one agent that can write to a business system or send information outside the company. On one page, record its owner, the identity it runs under, what it can read or change, what needs human approval, where its actions are logged, and how to stop it. Ask the owner to show you the settings and the stop test result.
Three checks you can run
| Check | What a pass looks like | What a fail looks like |
|---|---|---|
| Can the owner show the agent's permissions in each connected system? | Settings shown in each system's admin console, matching the job | A description from the builder or the vendor |
| Does the agent run under an identity of its own? | A dedicated identity with only the access the job needs | A person's account, or a shared admin login |
| Has someone paused the agent in a test and confirmed it could no longer use its connections? | A dated stop test, with the logs checked | No test, or nobody knows who would do it |
Questions we hear
Won't least privilege slow the project down?
It adds a short step before go-live, and it usually saves time later. An agent with limited, documented access is quicker to approve, easier to troubleshoot and simpler to expand. Start with read-only access, then add specific permissions as the workflow proves it needs them.
What if the platform's pause button doesn't cut off the agent's connections?
Then pausing isn't enough on its own. Write down which credentials or integrations the agent uses, who can revoke each one, and what else depends on them. Test revoking them in a pilot, so the team isn't learning the steps during an incident.
Which OWASP list applies to AI agents?
The OWASP Top 10 for Agentic Applications, published December 9, 2025, covers risks specific to agents. It is separate from OWASP's Top 10 for LLM Applications. The excessive agency guidance in the LLM list is still useful, because it describes the same least privilege and human approval controls.
Do we need a new security tool to run these checks?
Usually not. Permissions live in the admin consoles of the systems the agent connects to, and the stop test uses the agent platform's own controls and logs. The checks are a process: who owns the agent, what it can do, and how it stops.
Sources
- OWASP GenAI Security Project, Top 10 for Agentic Applications for 2026 (opens in a new tab)
- OWASP, announcement of the Top 10 for Agentic Applications (December 9, 2025) (opens in a new tab)
- OWASP, LLM06:2025 Excessive Agency (opens in a new tab)
- OWASP, Non-Human Identities Top 10 (2025) (opens in a new tab)
