Skip to content

AI Agent Permissions: What to Allow, Review and Block

Guides
A cobalt steel cabinet with one drawer open and a brass lock securing the drawer beside it.
Write down the AI agent permissions each job needs. An agent that sums up notes can read from one folder and make a private draft. Ask the draft’s owner before the agent saves to a shared file. Block the agent from deleting the source notes.

A sample AI agent permissions matrix

Start with three rules. Allow the steps that fall within the job. Review steps that have an effect beyond the draft. Block steps the job does not need. The table shows how this could work for a research and draft agent. Copy it and name your own folders, tools and people who can approve a change.
ActionRuleWhat the rule covers
ReadAllowRead only from the source folder named for the job.
DraftAllowMake a private draft for the person who asked.
EditReviewShow the owner the exact change to a named shared draft.
SendReview if neededShow the account owner the full message, address and attached files. Block sending for a draft-only agent.
BuyBlockKeep payment details and tools that make purchases out of the job.
DeleteBlockKeep the delete tool out. Reject each request to delete a source.
Change accessBlockUse the usual process for a change to roles or settings.

Set rules on the tools an agent can use

Use a read-only key for the source folder. Give the write tool access to the drafts folder alone. When someone asks the agent to act, check they can use the file or record in that request. Do the check where the tool runs, each time it runs. OWASP’s guide to excessive agency explains why AI agent permissions should match the work.
Think about what a person needs to see before they say yes. Show the name and current version of a file to be edited, along with the exact change. For an email, show who will get it, the text and any files it sends. n8n’s tool review guide shows how one tool can pause while it waits for a person to decide. Include any fixed address or other setting that affects the action.

When a request needs a new approval

Tie each approval to one request and its exact target. If the text, file version or address changes, ask again. Check that the action still fits the job. Block it if it does not. Keep a record of the request ID and the result so a retry does not do the work twice.
If the person says no or the request times out, stop the action. Ask a named backup if the work still needs review. Write down who asked, what tool would run, what it could change and who said yes or no. Keep private records and credentials out of the log. The Local AI guide explains why an agent can still act on an outside system when the model runs on your own machine.

Test AI agent permissions with sample data

Run a read, a write and a delete check before you give the agent real files. In our first check, the test read one sample note from the right folder. In the second, it asked before it wrote a shared draft. The third check blocked a delete request and kept the note in place.
We ran these checks on November 17, 2026, with a local Python test, throwaway files and simulated review choices. The test blocked changes to the text and the target after approval. It also stopped writes after a no or an expired request and blocked a repeat call. Run these checks with sample files in the tool you plan to use. Check the end result in the files as well as in the log.
loading

Loading