Is it safe to let an AI coding agent touch your computer?
By the Olomni team · Updated
It can be reasonably safe if you don’t leave it alone with anything that can’t be undone. An AI coding agent acts with your permissions: it can delete folders, publish your app, touch keys or run programs it downloads. What lowers the risk is being asked about the serious stuff in everyday words, having a copy of your files to go back to, and you deciding rather than the AI. No tool makes everything safe: Olo holds back and explains what it recognizes as sensitive and keeps that copy, but it is not an antivirus.
What can go wrong
- Deleting files or whole folders.
- Deleting or changing data in your database.
- Publishing changes to your real app or replacing what is already online.
- Reading or sending keys and passwords.
- Running programs it downloads.
- Changing who has access to your accounts.
Why approving blindly is the real problem
AI permissions are written for experts: “Allow rm -rf?”. When you can’t tell what it means, the easy move is to approve everything or skip permissions with --dangerously-skip-permissions, and then nobody reviews what gets deleted or published.
What lowers the risk
- Permissions in plain words: what the AI wants to do, what happens if you say yes and whether it can be undone.
- One more gesture for the irreversible: in Olo, holding on your phone.
- A copy of your files before what you approve (in Git projects).
- Categories that never happen on their own: moving money, trading stocks and changing access permissions.
- Checking “done”: seeing whether a test really passed after the last change.
What no tool covers, Olo included
An AI with access to your computer can do things Olo doesn’t recognize. And what already left your computer — a published change or data deleted on a service — can’t be undone from Olo. That is why Olo always tells you whether something can be undone.
Frequently asked questions
Is Olo an antivirus?
No. Olo warns about and holds back what it recognizes as sensitive and keeps a copy of your files so you can go back, but it is not an antivirus or full isolation.
Is it enough to ask the AI to be careful?
No. An instruction to the AI is not a barrier: the decision about anything irreversible has to be made outside it, by a person or by a rule.
What happens to my keys and passwords?
Olo asks before an AI reads a private key or sends something that looks like a secret, and masks keys and tokens in what it uploads to your account.