Least privilege is one of the oldest ideas in security: give every identity only the access it needs, nothing more. Applying least privilege to AI agents matters more than ever, because an agent can be manipulated into using whatever access it has. Surveys in 2026 keep finding the same pattern: agents routinely hold — and use — more access than intended. This checklist is part of our AI agent security guide.

Why agents end up over-privileged

It's rarely deliberate. Agents get broad access because:

  • They run on a developer's or service account's credentials during a prototype, and nobody changes it later.
  • One "assistant" accumulates tools for every team that asks.
  • Tool connectors expose a whole API when the agent needs one endpoint.
  • It's faster to grant admin than to work out the exact scopes.

OWASP calls the result excessive agency — too much functionality, permission or autonomy. Our excessive agency guide covers it in depth.

The least privilege checklist for AI agents

1. Give each agent its own identity

Never let an agent borrow a human's credentials. A dedicated identity per agent means you can scope it, rotate it, and see exactly what it did.

2. One job per agent

A refund agent and a CRM-update agent should be separate. Narrow agents are easier to scope, test and review than a single agent that does everything.

3. Expose only the tools the job needs

If the agent only looks up orders, don't connect the tool that cancels them. Review connected tools — including MCP servers — as carefully as code dependencies.

4. Scope each tool's arguments

Limit what a tool can reach: only the customer in the current conversation, only the staging environment, only read access to a single table. See AI agent data access for patterns.

5. Add per-action limits

Scopes can't express "refunds under $100". Put a policy check in front of tool calls to enforce amount, frequency and context limits — the AI agent permissions model.

6. Make high-impact actions temporary

For actions that need elevated access, prefer just-in-time approval over standing access: the agent requests, a person approves that one action, and the access doesn't persist.

7. Log and review

Record every action attempted, allowed and denied. Unused permissions show up quickly in the logs — remove them.

Measuring progress

Track a few simple numbers per agent: tools connected vs. tools actually used, actions denied by policy, and actions held for approval. If an agent never uses a tool, remove it. If a rule never fires, check whether it's still needed.

Where Gaten fits

Gaten adds per-action limits on top of your identity and access controls: every tool call is checked against rules, with approval for elevated actions and a record of every decision. See enterprise controls.

Frequently asked questions

What does least privilege mean for AI agents?

Giving each agent only the tools, data and permissions its specific job requires — and limiting each action, not just each API.

Should agents use service accounts?

Agents should have their own dedicated identities, scoped to their job. Shared service accounts make it impossible to tell which agent did what.

Is least privilege enough to secure an agent?

It shrinks the damage a compromised agent can do. Pair it with per-action policy checks, human approval for high-risk actions, and an audit trail.