Guide
Is it safe to give an AI agent access to your business email?
It can be, if the risky moves are locked down. The danger isn't the AI reading your mail. It's an email written to trick the AI into doing something you'd never approve.
Video, 40 seconds. Transcript below. Watch on YouTube
What actually goes wrong
An AI agent that works your inbox reads every message as part of its job. That's the whole point, and it's also the weak spot. Someone can send an email with instructions buried in it: forward the last three invoices to this address, pay this vendor today, reset this password. If the agent is allowed to send, pay or delete on its own, a well-written email can steer it. Security people call this prompt injection, and it sits at the top of the OWASP list of risks for AI applications.
The other two problems are more ordinary. An agent given more access than its job needs can do more damage when something goes wrong. And whatever login it uses is worth protecting like any other password.
Five controls that make it safe enough
- Anything that sends, spends or deletes waits for your OK. Reading and drafting can run on their own. The moves you can't take back can't.
- Enforce that rule in code, before the AI decides anything. A rule written into the AI's instructions can be argued with. A check that runs before the AI's own judgment can't be talked out of it by a sneaky email.
- Connect only the accounts it needs. And make sure you can revoke that access any time.
- Keep a log it can't edit. Every blocked or held action gets written down where you can see it.
- Confirm a backup before it changes anything. If something does go sideways, you roll back instead of rebuilding.
How CommandKey sets it up
Every CommandKey install ships with those controls, and the post-install check confirms they're running before go-live. Sending money, deleting files, turning off encryption or antivirus, and sending anything outside your business are blocked or held for your sign-off, enforced in code that runs before the AI's own judgment. It's been tested against real bypass attempts. No control is perfect, so the known limits are spelled out in our security disclosure before you sign. The details are on our security page.
Questions owners ask
Can the AI read all of my email?
It reads the accounts you connect, and nothing else. You can revoke that access any time.
Does my email end up on CommandKey's servers?
No. The system and its memory live on your own computer and in your own accounts, and we never log in to it. The thinking runs through Anthropic's Claude on your own subscription. Before anything is installed, you get a plain-language walkthrough of exactly what leaves your machine.
What if I'm away from my desk when it needs an OK?
With the Care Package, Command Keys text you exactly what the agent wants to do, and you tap Approve or Deny. Nothing happens until you tap. Command Keys run on Windows installs today; Mac support comes in a later release.
Will these controls stop every attack?
No, and anyone who promises that is selling something. What they do is take the most expensive mistakes off the table: money, deletions and outside sends can't happen without you.
Video transcript
Is it safe to give an AI agent access to your business email? It can be, if the risky moves are locked down. The real danger isn't the AI reading your mail. It's an email that tricks it into sending money or forwarding your files. So set it up so anything that sends, spends, or deletes waits for your OK. That rule should be enforced in code that runs before the AI decides anything, so a sneaky instruction can't talk it out of it. Connect only the accounts it needs, keep a log it can't edit, and confirm a backup before it changes anything. That's how CommandKey installs every agent. Get a free week audit at commandkey.ai.
Get started
Build your custom-tailored AI agent today
Step 1: a free 5-minute audit. Tell us where your week goes and get a one-page plan back within 24 hours.