Least privilege for AI agents

Least privilege means an AI agent gets only the tools, data and permissions its task needs, for only as long as it needs them. It limits what a mistake or an injected instruction can do.

Why it matters

An agent acts with the permissions it was given, not the ones its task needed. If it holds an admin token, an injected instruction holds that token too. Most serious agent failures are ordinary access used for the wrong end: see excessive agency.

In practice

  • Fewer tools. Offer the tools the task uses. A support agent does not need a shell.
  • Narrow scopes. Read-only where reading is enough. One repository, one table, one mailbox.
  • Short-lived credentials. Issue them per task and let them expire.
  • The user's own rights. An agent acting for a user should reach what that user can reach, not what the service account can.
  • Limits. Cap spend, calls and run time.

Where permissions stop

Permissions say what an agent can do. They cannot say whether a permitted action is right this time: an agent allowed to send email can still send the wrong file to the wrong person. That takes a check on the action itself, with its arguments, before it runs. See tool call guardrails.

With White Circle

Control reads each tool call and its arguments against your policies before it runs, so a permitted tool used for the wrong end is blocked or held for approval.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL