What is least privilege access?
Least privilege access is a security principle that gives each user, application and AI agent only the access it needs to do its current job, and nothing more. It limits the damage from a mistake, a stolen credential or a misused tool, because each identity can reach only a small part of the data.
For example, a support rep needs a customer’s order history to answer a shipping question. Under least privilege, the rep’s tools show the orders and shipping address but not the card number or the customer’s other accounts. If the rep’s login is stolen in a phishing attack, the attacker gets the same narrow view.
How least privilege is applied
- Start from no access and add only what each job needs.
- Grant extra access for a limited time when a task needs it, then remove it (often called just-in-time access).
- Review access regularly and remove rights that are no longer used.
- Update access when roles change, so people do not collect rights as they move between teams.
- Apply it to machines too, including service accounts, API keys and AI agents.
Least privilege appears as control AC-6 in NIST SP 800-53 and is a core idea in zero trust architectures.
Least privilege and related ideas
| Idea | What it means |
|---|---|
| Least privilege | Each identity gets the minimum access for its job |
| Need to know | Access to specific information only when the work requires it |
| Separation of duties | No single person controls every step of a sensitive process |
| Zero trust | Every request is verified, with no trust based on network location |
Why least privilege is harder with AI agents
AI agents are often given broad service accounts so they can handle any task a user might ask for. That turns one AI agent into a single point that can reach many systems. A misleading instruction or a wrong tool call can then pull data far outside the task.
Data that reaches a model also spreads. It lands in the context window, logs, AI agent memory and sometimes a vector store. So least privilege for AI has two parts: limit which systems an AI agent can call, and limit which real values inside the data reach the model for each task.
The second part matters because AI agents often need some of a record but not all of it. An AI agent drafting a shipping update needs the order number and city, not the card number or date of birth. Keeping those values masked lets the task finish while the extra data stays out of reach.
How Protecto helps
Protecto applies least privilege to the data itself. Sensitive values are tokenized before they reach models and AI agents, and real values are revealed only to the users and tools the policy names for that task. Every access decision is logged, which gives auditors a record of who saw what. See data access control for AI agents.
Frequently asked questions
What is an example of least privilege?
A support rep can see a customer’s orders and shipping address but not the full card number, because answering shipping questions does not need it.
How does least privilege relate to zero trust?
Zero trust verifies every request. Least privilege decides how much access a verified request gets. Zero trust architectures rely on both.
How do you apply least privilege to AI agents?
Limit the tools and systems each AI agent can call, and control which sensitive values reach the model for each task, so other values stay masked. Read more in the principle of least privilege.
Related terms
Role-based access control (RBAC) · Attribute-based access control (ABAC) · Context-based access control (CBAC) · Data minimization · AI agent