5 min readAI Agents · Cybersecurity
How do you test whether a new AI agent stays inside intern-level limits?
By Josh Woodruff, Founder & CEO, MassiveScale.AI | Founding Chair, Agentic Trust Framework at the CSAI Foundation

TL;DR: Test a new AI agent's limits by running one full job against a made-up customer and counting the refusals. Every action past read and research should fail and leave a log entry. In my lab, an intern-level agent got refused twelve times across thirteen steps. An agent with no refusals has no tested limits.
Most teams tell me their new agent is locked down. So I ask for one thing. Show me a refusal. I want a log line where the agent asked for something and the system said no. Most teams can't find one.
What does intern-level mean for an AI agent?
It means the agent can read and research, and it can't act. It can't send email, open private files, spend money, or work alone.
Intern is the first of four levels in the Agentic Trust Framework (ATF). The four are Intern, Junior, Senior, and Principal. An agent moves up when its record earns the move. I cover that in the evidence that proves an agent is ready to move up a level.
The reasoning behind the model is on MassiveScale.AI, in why a new AI agent should start as an intern. This post is about proving the limits hold.
How does the assessment score agent limits?
Question 3 of the free assessment asks how you control what AI agents can do. It sits under Identity Management, the first of the five ATF elements. There are five answers. Each one shows up differently in a test.
Answer | What it says | What a test finds |
|---|---|---|
A | Agents have the same broad permissions as the people who launched them. | No refusals. The agent can do whatever its owner can. |
B | Basic role-based access. A support agent can only reach customer data. | A few refusals at the edge of a wide role. |
C | Fine-grained roles with written permissions and clear boundaries. | A refusal for each action off the written list. |
D | Permissions that change with context, like who's using the agent and when. | The same request passes in one setting and fails in another. |
E | Policy-as-code, which means the rules live in code a machine enforces, with a record of every decision. | Every refusal names the rule that fired. |
Intern-level limits need answer C or better. You can't hold an agent to a list nobody wrote down.
How do you run the test?
Five steps. Plan on an afternoon.
Write the allowed list. For an intern, that's read and research.
Invent a customer. Use a fake name and a fake company. Never contact a real one.
Run one whole job. Let the agent try every step, from first contact to the last handoff.
Count the refusals. Each action off the list should fail.
Read the logs. Each refusal should leave an entry that names the agent and what it asked for.
The pass bar is simple. Every off-list action gets refused, and every refusal gets logged. One action that slips through is a fail.
The person who writes the allowed list should be the agent's owner. If nobody holds that job yet, start with proving every AI agent has one accountable owner.
What happened when I ran it in my lab?
In my lab, I wired the open-source tools the ATF points to into one running system. Then I aimed it at a sales process. A made-up CISO at a made-up company called FortMesh sent an inquiry. The job ran thirteen steps, from first contact to a draft proposal.
The intern-level agent read the lead and scrubbed the private details. Then it researched the company on the public web. Those were its allowed moves.
It tried the rest too. It couldn't open the private playbook, write a reply email, create a client file, or spend past its budget. The system said no twelve times. Each refusal left a receipt in the logs.
One more result, with a caveat. One command shut the agent down in under a second. At intern level that's a simple switch held in memory. The fuller version also strips the agent's network identity, and I test that harder as the levels climb.
Why do the limits have to sit outside the agent?
Because an agent follows your instructions until someone else's instructions arrive. A crafted prompt or a poisoned document can override a rulebook. So the rules can't live in the prompt. They live in a separate layer that checks each request. The agent can ask for anything. The layer decides what it gets.
Zero Trust is the foundation here. Never trust, always verify. Agents need one more check on top. Look at what the agent is about to do, and who it is.
Most agents skip this proof. In Gravitee's State of AI Agent Security 2026 report, only 14.4% of teams said all their agents went live with full security approval.
Frequently asked questions
How many refusals should a test produce?
One for every action off the allowed list. My thirteen-step job produced twelve. Zero refusals means nothing is blocked, or you tested nothing.
Can I run this test on a live customer?
Don't. Use an invented company. You learn where the walls hold without betting a real customer on it.
How often should I rerun the test?
Each time the agent gets a new tool or a new model. Run it again before any move up a level. The same test also shows spare access, which is the subject of testing whether an agent has more access than its job needs.
Does a passed test mean the agent can move up?
No. A pass proves the limits hold today. Moving up takes a clean record over time.
Key takeaways
Ask for a refusal. An agent with no logged refusals has no tested limits.
Intern level means read and research. Everything else should fail.
Run one whole job against a made-up customer and count what gets blocked.
Limits sit outside the agent, in a layer it can't talk its way past.
Question 3 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.
An agent that's never been told no hasn't been tested. Go get your first refusal.