← sources

Least privilege still applies when the "user" is an AI agent

source: reddit (reddit:sysadmin)published 2026-10-19 20:01ingested 2026-10-20 02:39

retained text

We're all getting pitched AI agents that can read tickets, query systems, send email and make config changes on their own. Useful, but an agent is just another identity with access. The difference is that it acts in seconds, doesn't get tired, and does exactly what it's told, including by a malicious actor who deceives it.

The basics we already apply to people and service accounts still apply here:

* **Give it its own account.** Never let it run under your admin credentials or a shared login.
* **Scope access to the job.** If it summarizes tickets, it gets read-only access to tickets, not the whole tenant.
* **Separate read from write.** Anything that sends, deletes or changes config should need a human to approve it, at least until you trust it.
* **Use short-lived tokens with an expiry,** not a forever API key sitting in a config file.
* **Log everything it does** somewhere it can't edit, and actually review it.
* **Have a kill switch.** Know how to revoke its access in one step.

A person with too much access might make one bad change before someone notices. An agent with too much access can make a hundred before you finish your coffee.

How are you handling permissions for the AI tools creeping into your environment? 

document fetched 2026-10-20 00:36 · canonical https://reddit.com/r/sysadmin/comments/1xa7e0q/least_privilege_still_applies_when_the_user_is_an