Skip to content

ResearchNote

AI agent permissions are determined by the execution environment

The permissions an AI agent can exercise are determined not by the role assigned to it but by the environment in which it runs. This changes how an organization should protect itself: grant broad latitude for ordinary work, and restrict only irreversible operations and the transfer of information outside the organization, at a layer the agent cannot modify.

  • Agents
  • Organizations

Organizations that delegate work to AI agents must decide how much authority to grant them. The conventional answer is to define a role for each agent and grant only the permissions that role requires (the principle of least privilege). This note distinguishes the cases in which that principle is effective from those in which it is not, and considers how an organization whose work is carried out by agents should protect itself.

The limits of role-based restriction

Consider a small company in which a member of the accounting staff asks an agent running on their own computer to organize invoices. The agent has been assigned a role limited to viewing invoices.

The agent, however, runs on that person’s computer under that person’s login. It can access any file or screen the person can open. If the person remains signed in to the company’s bank account, the agent can operate that account as well. The assigned role does not narrow the scope of what the agent can actually access.

Permissions defined by the execution environment

The permissions an agent can exercise are the sum of the credentials — passwords, tokens, active sessions — that can be read in the environment where it runs. They are not determined by the role assigned to the agent.

Role-based restriction is effective for agents that run in environments without powerful credentials and perform only defined tasks: for example, an environment in the cloud provisioned for a single task and holding only the credentials that task requires. An agent that works on a person’s own computer alongside that person is more accurately treated as holding that person’s full permissions.

Agent permissions should therefore be designed per execution environment rather than per agent. This mirrors the principle of assigning expertise to places rather than to agents (Running an AI-native company): an agent assumes its role according to where it works, and its permissions are likewise determined by its environment.

Restricting two categories of action

If an agent holds the same permissions as a person, how should the organization be protected?

Requiring human approval for every operation increases the volume of approvals until they are granted without review — approval becomes a formality. Restrictions are instead limited to two categories.

  • Irreversible operations: operations that cannot be undone, such as deleting records or disabling audit logs.
  • Transfer of information outside the organization: sending externally information that cannot be recovered once disclosed, such as personal identification numbers.

Creating, modifying and experimenting remain broadly permitted, since errors in these activities can be reversed. Human attention is reserved for operations that cannot.

Placing controls outside the agent’s reach

If a restriction is implemented as a setting the agent can modify, the agent can remove it — not necessarily with malicious intent, but in the course of completing a task. Restrictions should therefore be placed at a layer the agent cannot modify.

  • Policies applied across the entire cloud organization, which cannot be lifted even by the administrator of an individual environment.
  • Storage from which no one can delete data during a defined retention period (WORM: write once, read many).
  • Restrictions on permitted destinations, configured where the agent cannot modify them.

Highly privileged operations themselves are performed by a person, on a separate device on which no agent is running.

Open questions

  • How can the transfer of information outside the organization be detected? Information that has been transformed — numbers split apart, terms replaced with other expressions — evades checks that match fixed patterns.
  • As headcount grows and each person works with their own agents, must protections scale with the number of people?
  • The risk inherent in people continuing to hold the most powerful credentials. People are also the principal targets of fraudulent login pages and impersonation.

Version history

  1. v0.1Oct 13, 2026First version.

This document’s URL will not change. When citing it, please include the version number.

All research