Home / Glossary / Least privilege for AI agents
Glossary
Least privilege for AI agents
Least privilege for AI agents is the practice of giving each agent only the access its task needs. It applies an old security principle to software that acts on its own. An agent that needs one code repository should not also hold the keys to payroll or production.
By the AIBound team · Updated 17 November 2026
Key takeaways
- Least privilege gives each AI agent only the access its task needs.
- An agent on a laptop usually starts with all of its employee's access, which is far more than one task needs.
- Limits work at two levels: what the agent can reach, and which actions it may take.
Why least privilege matters for AI agents
Least privilege is a long-standing rule. The NIST glossary describes it as giving each entity the minimum resources and authorizations it needs to do its job. For people, companies apply it with roles and access reviews.
AI agents break the habit. An agent on a laptop runs with the employee's logins, tokens and keys, so it starts with everything the employee has. If a hidden instruction tricks the agent, the damage is as wide as that access. Narrow access means a mistake or an attack stays small.
How to apply least privilege to AI agents
Least privilege for agents starts with a map of what each agent can reach today. The limits come after the map.
- 1 List each agent, the account it uses and the systems it is wired into.
- 2 Compare that reach with the task the agent is there to do.
- 3 Remove the connectors, tokens and MCP servers the task does not need.
- 4 Set rules for the rest, so risky actions are asked about or denied.
Least privilege for AI agents vs least privilege for users
The principle is the same. The way it is applied differs, because an agent borrows a person's access and acts much faster.
| Question | Least privilege for AI agents | Least privilege for users |
|---|---|---|
| Where does the access come from? | The employee's account, tokens and connectors | Roles and groups set in the directory |
| How is it limited? | Fewer connectors, plus rules on each action | Narrower roles and regular access reviews |
| Who reviews it? | Often nobody, because agents are on no access list | Managers and the identity team |
Sources
- NIST glossary: least privilege · csrc.nist.gov
Sources checked 17 November 2026
Examples
Least privilege examples for AI agents
FAQ
Frequently asked questions
What is least privilege for AI agents?
Least privilege for AI agents means each agent gets only the access its task needs. An agent that writes code needs its repository. It does not need payroll, HR data or production keys. Narrow access limits the harm if the agent makes a mistake or follows a hidden instruction.
Why are AI agents often over-permissioned?
Most agents on a laptop run as the employee who started them. They pick up that person's logins, tokens and keys without anyone choosing to grant them. Each new connector or MCP server adds more reach. Agents also sit outside the access reviews that catch the same problem in user accounts.
How do you enforce least privilege for AI agents?
Work at two levels. First, cut what the agent can reach by removing connectors, tokens and MCP servers the task does not need. Second, set rules on actions, so a risky command or file read is asked about or denied. The rules should run where the agent runs.
How does AIBound help with least privilege for AI agents?
AgentTracer maps who or what each agent runs as and what it is wired into, and it surfaces agents with more reach than their job needs. AgentHarness then enforces plain-English policy inside the agent, so each action is allowed, asked about or denied in real time.