THE AGENT PERMISSIONS REVIEW
Understanding AI HQ free worksheet, version 1.0 (2026-11-12)
Fill it in online, print it or get the PDF: Understanding AI HQ website, Learn > Worksheets

HOW TO USE IT
1. Fill it in from the tool's own documentation, before you connect anything.
2. List every action, and mark which ones change things or can't be undone.
3. Review it again whenever the tool updates or you add a connection.

================================================================
1. WHAT I'M CONNECTING
================================================================

AGENT, MCP SERVER OR SKILL
>

WHO MADE IT, AND WHY I TRUST THEM
>

WORKFLOW OR AGENT?
[ ] Fixed steps I set (workflow)
[ ] AI chooses the steps (agent)

================================================================
2. ACTIONS
================================================================
Allow: runs without asking. Confirm: asks me each time. Off: switched off.

WHAT IT CAN DO
          |  Action  |  Read-only, changes things, or can't be undone  |  Allow, confirm or off
--------  |  ------  |  --------------------------------------------  |  ---------------------
Action 1  |  ____    |  ____                                          |  ____
Action 2  |  ____    |  ____                                          |  ____
Action 3  |  ____    |  ____                                          |  ____
Action 4  |  ____    |  ____                                          |  ____

================================================================
3. OUTSIDE CONTENT AND INJECTION
================================================================

OUTSIDE CONTENT IT WILL READ (WEB PAGES, EMAILS, FILES FROM OTHERS)
>
>

PROTECTIONS
[ ] Nothing that sends, deletes, publishes or spends runs without my confirmation
[ ] It has the narrowest access that does the job
[ ] I'll read each confirmation prompt, not click through
[ ] I know how to disconnect it

----------------------------------------------------------------
Limit what an agent can do, not just what it's told. Anything it can't do, injected text can't make it do.


################################################################
# EXAMPLE ANSWERS
################################################################

THE AGENT PERMISSIONS REVIEW
Understanding AI HQ free worksheet, version 1.0 (2026-11-12)
Fill it in online, print it or get the PDF: Understanding AI HQ website, Learn > Worksheets

HOW TO USE IT
1. Fill it in from the tool's own documentation, before you connect anything.
2. List every action, and mark which ones change things or can't be undone.
3. Review it again whenever the tool updates or you add a connection.

EXAMPLE ANSWERS. The example is illustrative, based on the email scenario in the lesson. It isn't a review of any real product.

================================================================
1. WHAT I'M CONNECTING
================================================================

AGENT, MCP SERVER OR SKILL
> An email MCP server for my AI desktop app

WHO MADE IT, AND WHY I TRUST THEM
> The email provider's own official server, linked from its documentation.

WORKFLOW OR AGENT?
[ ] Fixed steps I set (workflow)
[x] AI chooses the steps (agent)

================================================================
2. ACTIONS
================================================================
Allow: runs without asking. Confirm: asks me each time. Off: switched off.

WHAT IT CAN DO
          |  Action                  |  Read-only, changes things, or can't be undone  |  Allow, confirm or off
--------  |  ----------------------  |  --------------------------------------------  |  ---------------------
Action 1  |  Search and read emails  |  Read-only                                     |  Allow
Action 2  |  Create a draft          |  Changes things                                |  Allow
Action 3  |  Send an email           |  Can't be undone                               |  Off
Action 4  |  Delete an email         |  Can't be undone                               |  Off

================================================================
3. OUTSIDE CONTENT AND INJECTION
================================================================

OUTSIDE CONTENT IT WILL READ (WEB PAGES, EMAILS, FILES FROM OTHERS)
> Every incoming email, including from strangers.

PROTECTIONS
[x] Nothing that sends, deletes, publishes or spends runs without my confirmation
[x] It has the narrowest access that does the job
[x] I'll read each confirmation prompt, not click through
[x] I know how to disconnect it

----------------------------------------------------------------
Limit what an agent can do, not just what it's told. Anything it can't do, injected text can't make it do.