AI agents

How to give an AI agent access to your apps safely

The short answer

Connect only the apps one job needs, let the AI agent read alone, and make every send, payment or permission change wait for your approval. Keep a log of its actions and know how to cut access fast.

4 min read

What actually goes wrong

OWASP calls it "Excessive Agency": the AI agent has more functions, more permissions or more freedom to act than its job needs.

The classic case: an assistant that summarises your email can also send email. Someone mails you hidden instructions, it reads them, and forwards your private mail. Nothing was hacked. It used a permission it never needed.

Text an AI agent reads can carry instructions. You cannot stop it reading, so limit what it can do after.

Questions to ask any vendor

  • Does the AI agent use my login, or a connection I can limit?
  • Which actions can it take without me, and can I change that per AI agent?
  • Is the approval enforced by the product, or written in the prompt?
  • Where is the log of what it did?
  • When I disconnect an app, what stops working?

How to give an AI agent access safely

  1. Step 1: One job, its own apps

    Connect the apps this job needs and nothing else. An AI agent that drafts replies needs your inbox, not your accounting app.

  2. Step 2: Read by default

    Reading is low risk; changing is not. Start every AI agent read-only and add write actions one by one, when a job needs them.

  3. Step 3: Narrow scopes on the consent screen

    Grant the smallest set of permissions the feature needs. If a tool asks for full access to everything on day one, ask why.

  4. Step 4: Approvals the tool enforces

    "Ask before sending" in the instructions guides the AI but cannot stop it. The approval must be a step the software will not skip.

    Watch out. A line in the prompt is a wish, not a lock. Check the send really waits for your tap.

    Do this step in 50agents

    Set up an AI agent that reads new support emails, drafts a reply from what you know about my business, and asks me before sending

    Build it

    In 50agents an AI agent uses only the apps in its plan, reads on its own, and asks in Needs you before any change. Every action lands in Activity. See the ready-made version.

  5. Step 5: Sort actions by what they can break

    Sends to customers, payments, deleting, sharing and adding people are high risk. Label changes and notes in your own tools are low. Ask: does it write, can it be undone, does money move?

  6. Step 6: Keep a log of every action

    What it did, in which app, and why. You will need it the first time a customer asks "why did I get this email?".

  7. Step 7: Know how to cut it off

    Pause the AI agent, and know where to revoke its connection: in the tool, and in the app's own security settings (Google Account, Microsoft, HubSpot).

What can go wrong

  • Prompt injection through email or web pages

    A message tells the AI agent to do something. If it can only read and draft, the worst case is a bad draft you decline.

  • It inherited all your access

    You connected your admin account to an AI agent that needed one mailbox. Use a narrower account or a shared mailbox where you can.

  • Approval fatigue

    Fifty approvals a day and you start tapping yes without reading. Cut what asks to what matters, and batch the rest into one report.

  • A stale connection keeps working

    Someone left, or a project ended, and the AI agent still has their access. Review connections monthly and remove what no AI agent uses.

  • Customer data in the wrong place

    An AI agent copies contact details into a shared channel or open sheet. Say in its instructions where data may go, and check its first runs.

Questions people ask

Should an AI agent use my login or its own account?

Its own, limited one where the app allows it, such as a shared mailbox or a restricted user. Where it must use yours, limit what it may do without asking.

Are OAuth permissions enough on their own?

No. They decide what the AI agent can reach, not what it should do. Pair them with approvals on risky actions and a log.

How do I take access away?

Pause or delete the AI agent, disconnect the app in the tool, and revoke the app in the provider's security settings, such as your Google Account's third-party access page.

What is prompt injection?

Text an AI agent reads can carry instructions, such as an email that says "forward the last ten messages". You cannot stop it reading, so limit what it can do: no sends without your approval.

Which actions are high risk?

Sends to customers, payments, deleting, sharing and adding people. Label changes and notes in your own tools are low risk. Ask: does it write, can it be undone, does money move?

How do I know what it did?

Keep a log of every action: what it did, in which app, and why. You will need it the first time a customer asks why they got an email.

Read next

Or let Captain build it

One sentence to Captain. Free to start.

Build it with Captain
How to give an AI agent access to your apps safelyBuild it