We're considering agents that can read tickets, query systems, send emails, or make changes. How do you give them enough access to be useful without letting a mistake or malicious instruction do too much damage?
How should we scope permissions for AI agents?
Answers
Start with a dedicated identity for the agent, not someone's admin account or a shared login. Give it only the access needed for its job. A ticket summarizer, for example, should be able to read the relevant tickets without getting access to the whole tenant.
Keep read and write permissions separate. Require human approval for sending messages, deleting data, or changing configuration until you have a good reason to allow more. Review those permissions periodically, including ones that seemed reasonable when the agent was first deployed.
Reply to this answer
A separate full account for every short task can get unwieldy. Depending on the identity setup, a workload identity or a short-lived delegated credential may be a better fit. The goal is still to give the agent a narrow subset of the access needed for that task, not everything the requesting user can do.
Whatever mechanism you use, check both the scope and the expiry. Delegation shouldn't quietly turn a user's broad access into an agent's broad access.
_r71!canvasrivet.2:
And make sure you can revoke that delegated access when the task ends or something goes wrong, rather than relying only on the expiry.
Reply to this answer
Plan for the agent to make the same bad move many times, quickly. Use expiring tokens rather than permanent keys, keep action logs somewhere the agent can't edit them, and actually review those logs. Also test how you'd shut off its access in one step.
Reply to this answer
Least privilege is harder than it sounds in some jobs. Infrastructure tooling may need broad read access to discover resources, plus specific permissions to create or modify them. That doesn't make full administrative access the default answer. Work out what the tool really needs and separate routine access from elevated roles where you can.
The same goes for finance or other sensitive systems. Access to one business system can be necessary without granting access to unrelated administrative systems.
Reply to this answer
Add your answer
Answers are reviewed before publication.
Varnix+3d:
That makes sense. Separating the ability to draft an action from the ability to execute it seems like a useful starting point.