5 min readAI Agents · Privilege Changes for AI Agents
How do you test whether an AI agent has more access than its job needs?
By Josh Woodruff, Founder & CEO, MassiveScale.AI | Founding Chair, Agentic Trust Framework at the CSAI Foundation

TL;DR: Compare what each AI agent was granted with what it used in the last 30 days. Any grant with no use and no line in the agent's job description is spare access. Remove it. The pass bar is zero spare grants and one identity per agent. Expect to find spare access on the first run.
Nobody decides to give an agent too much access. It piles up. A permission gets added for one project and never comes off. A year later, no one can say why the agent reaches what it reaches. A test finds the pile.
What counts as too much access for an AI agent?
Anything the agent can reach that its job doesn't need. A support agent needs to read account details. It doesn't need to reverse charges.
Security people call the rule least privilege. Each agent gets the smallest access that still lets it do its job.
Spare access is risky for a plain reason. An agent works toward its goal at machine speed, and spare access is one more way to get there. No attacker is required. The background is on MassiveScale.AI, in how to give an AI agent the right amount of access.
How does the assessment score agent access?
Question 3 of the free assessment asks how you control what AI agents can do. It's part of Identity Management, the first of the five ATF elements. Each of the five answers leaves a different mark when you compare grants with use.
Answer | What it says | What the access diff shows |
|---|---|---|
A | Agents have the same broad permissions as the people who launched them. | The agent's grant list is its owner's grant list. Most of it goes unused. |
B | Basic role-based access. A support agent can only reach customer data. | A shorter list that still holds grants the job never touches. |
C | Fine-grained roles with written permissions and clear boundaries. | Grants match a written list, with a few stale ones. |
D | Permissions that change with context, like who's using the agent and when. | Grants appear for a task and drop away after it. |
E | Policy-as-code, which means the rules live in code a machine enforces, with a record of every decision. | Each grant traces to a rule, and each use traces to a log entry. |
How do you run the access diff?
The diff is a subtraction. Granted minus used.
Write the job in one sentence. Name the systems the job needs.
Pull the granted list. Include every scope and every key the agent holds.
Pull the used list. Take it from 30 days of logs.
Subtract. Whatever was granted and never used goes on the spare list.
Check each spare grant against the job sentence. If the job doesn't name it, remove it.
The pass bar has two parts. The spare list is empty. Each agent runs under its own identity.
Step 3 can fail in a telling way. If the logs can't tell your agents apart, you've found a shared login, and that's the first thing to fix.
Why does a shared login break the test?
Because every action looks the same in the logs. When five agents run under one account, you can't build a used list for any one of them. You also can't shut off the one that's misbehaving without stopping the other four.
It's common. In Gravitee's State of AI Agent Security 2026 report, 45.6% of teams said they still rely on shared API keys between agents. Only 21.9% treat agents as identities of their own.
What do you do with each kind of spare access?
Spare access shows up in four patterns. Each has its own fix.
Pattern | What the diff shows | The fix |
|---|---|---|
Shared account | One used list for many agents | One identity per agent |
Inherited human access | The agent can reach everything its owner can | A role scoped to the agent's job |
Permanent keys | A key issued once that never expires | Keys that expire in hours |
Stale project grant | A grant with no use in 30 days | Remove it, and grant it again only for the task |
Permanent keys deserve their own pass. I cover that in scoring which rung of the credential ladder each agent is on.
Frequently asked questions
Why 30 days?
It's long enough to catch monthly jobs. If an agent runs a quarter-end task, use 90 days for that agent.
Won't removing access break something?
Sometimes. Remove one grant at a time and watch for refusals. A logged refusal tells you which grant to put back. The method is the same one I use in testing whether a new agent stays inside intern-level limits.
How often should I run the diff?
Every quarter. Run it again whenever the agent's job changes.
Should each AI agent have its own login?
Yes. Without one, the diff can't be run at all.
Key takeaways
Spare access is any grant with no use and no place in the job description.
The test is a subtraction: granted minus used, over 30 days.
The pass bar is an empty spare list and one identity per agent.
A shared login fails the test before it starts.
Question 3 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.
Spare access piles up on its own. It only comes off when someone runs the subtraction.