4 min readAI Agents · Privilege Changes for AI Agents
How do you test what an AI agent can reach in ten minutes?
By Michelle Savage, Co-author, Agentic AI + Zero Trust

TL;DR: Pick one AI agent and start a timer. Ask your team to list what it reads, what it writes, when its key expires, and who can shut it off. A full answer in under ten minutes is a pass. If the answer is "it uses the service account," or the clock runs out, nobody holds the list.
"What can this agent do?" is the question everyone asks. It describes a job.
"What can this agent reach?" is the better one. It describes how much of the business is exposed when the agent does something nobody planned.
And you can test the answer with a kitchen timer.
Why test reach and not the job?
Because the job is what you meant. The reach is what you built.
Teams scope the job and skip the access. An agent gets built to answer refund questions. Somewhere in setup it gets a key that opens far more than refunds. Nobody chose that. The key was sitting there, and it worked.
Leaders tend to feel fine about this. In Gravitee's State of AI Agent Security 2026 report, 82% of executives said they were confident their policies protect them from agent actions nobody approved. A timer is a quick way to check that confidence.
There's more on the question itself on MassiveScale.AI, in what your AI agent can access right now.
How does the assessment score this?
Question 19 of the free assessment asks how agent access to resources is controlled. It's part of Segmentation, one of the five ATF elements. Here's how each answer sounds when the timer is running.
Answer | What it says | What you hear in the ten minutes |
|---|---|---|
A | Agents can reach anything they need, with no restrictions. | "It can get to whatever it needs." |
B | Basic access restrictions are in place. | "It uses the service account." |
C | Each agent has a written list of the resources it's allowed to use. | Someone opens a document and reads you the list. |
D | The allowed lists are enforced, with monitoring and alerts. | Someone shows you the list and the last alert it raised. |
E | Policy-as-code, which means rules in code decide each request based on context. | Someone shows you the rule that decides, and its log. |
How do you run the ten-minute test?
You need one agent and the team that runs it. Bring a timer.
Pick the agent closest to your customers. That's where a bad afternoon gets expensive.
Start the timer.
Ask what it can read. You want system names. "Customer data" is too vague.
Ask what it can write or change.
Ask when its key expires and who can shut it off.
Stop the timer. Write down the time and the answers.
Here's how to grade it.
What you get | Verdict |
|---|---|
Two systems it reads, none it writes, a key that expires daily, and a named person with an off switch | Pass |
"It uses the service account" | Fail. Nobody knows the reach. |
A complete answer that took an hour | Fail. The list exists, but nobody holds it. |
No answer by the time the clock runs out | Fail. That's your finding. |
What should a full answer include?
Four things. Think of the agent as a contractor with a badge.
What it can read, by system.
What it can write or change, by system.
When its key expires by itself.
Who can shut it off today, by name.
You'd never hand a new hire a master key on day one. Most AI agents get less of a once-over than the summer intern, and they work all night.
What do you do when an agent fails?
Put everything it reads and writes on one page. That page is the list nobody had.
Then trim it. If the agent "needs everything," it's probably doing several jobs on one key. Split the jobs and scope each one.
A failed test also tells you how big a mistake could get. Josh covers that in measuring an AI agent's blast radius before the inventory is finished.
Frequently asked questions
Isn't this just normal access management?
Mostly, and that's good news. You already have the practice and the people. What's new is how many agents there are and how fast they move.
Who should answer the question?
The team that built the agent answers. Security checks the answer. If the builder can't produce the list, the agent isn't ready for production.
Does this apply to agents we bought from a vendor?
Yes, and those are often harder. Ask the vendor the same four things. If they can't answer, write that down as the result.
What if the agent reaches its tools through MCP?
Then each server needs its own check. Josh lays that out in testing whether your MCP servers are under control.
Key takeaways
Test reach. The job description won't tell you what's exposed.
One agent, one timer, four questions.
A full answer in under ten minutes is a pass.
"It uses the service account" is a fail.
Start with the agent closest to your customers.
Question 19 is one of 30 in the free assessment. It takes about ten minutes too, and it scores you on all five ATF elements.
The question costs nothing to ask. Not knowing the answer is the expensive part.