Skip to content
Cloud Least Privilege

Guide, 10 Nov 2026, read 10 Nov 2026

A boundary is a limit set above the grants, so one change can remove a class of access

See the rankingAll guides

A grant gives access and a boundary limits it

A grant says an identity may do something. A boundary says what no grant beneath it can allow. The two are evaluated together, and the effective result is what both permit. That is why a boundary can remove access without touching the grants: the broad role stays in place and the boundary makes parts of it inert. The same shape appears in every cloud, even though the names and the evaluation rules differ in each one.

This site uses the word in the general sense. The exact policy types, their evaluation rules and the traps in each are the subject of dedicated references outside this site, and this guide names one example per cloud only to show the shape. A reader should not expect the three examples below to behave alike, since each cloud evaluates its own policy types in its own order.

Each cloud documents its own version

AWS documents a permissions boundary as an advanced feature that uses a managed policy to set the maximum permissions an identity-based policy can grant to an IAM user or role. The entity can perform only the actions allowed by both its identity-based policies and its boundary. It is a ceiling on one identity and does not grant anything itself. AWS also lists permissions boundaries as a way to delegate permissions management within an account.

Google Cloud documents Principal Access Boundary policies as policies that define the resources principals can access. They attach to principal sets, in contrast to allow and deny policies, which attach to resources. Google gives the example of preventing principals from accessing resources in other organizations, which can help prevent phishing attacks or data exfiltration. Google notes that by default principals are eligible to access any Google Cloud resource, so the boundary narrows an otherwise open default.

Microsoft documents deny assignments in Azure RBAC as attaching a set of deny actions to a user, group or service principal at a scope, blocking those actions even if a role assignment grants them. The effect is a block that holds regardless of the grants beneath it, at the scope where it is attached. The page covers listing them, which is the first step in learning which denies already exist in your tenant.

Why a boundary removes a class and a clean-up removes an item

A permission-by-permission clean-up deletes one grant per change and has to be repeated for every identity. A boundary applied at an organization, folder or account level acts on every identity beneath it at once. If the boundary denies a service nobody uses, a region nobody runs in, or access from outside the organization, thousands of grants become irrelevant in one change. The comparison is between ten thousand edits and one reviewed policy.

This is what the criterion "One change, a class of risk" scores on this index, at 15 percent of the total. Sonrai says it applies global default deny in one action across an organization while used permissions stay available. Act says its guardrail campaigns use broad explicit-deny policies for whole attack categories, with built-in exceptions. Both are vendor claims; the method describes how they were scored.

A boundary needs a simulation more than a grant does

Because a boundary acts on many identities, its mistakes are wide. A deny that is slightly too broad blocks a deployment role in forty accounts at the same moment. The same reach that makes a boundary efficient makes an untested one dangerous. Release should be gradual for the same reason, starting with the accounts that have the fewest dependencies. Wide reach also means a wide review, since the list of affected principals can run to hundreds and each needs an owner.

That is why the index scores safe enforcement at 20 percent, the second-largest weight. The practical rule is to replay recent access under the proposed boundary, list every principal that would change outcome, and release it in audit mode first. How to simulate a permission change walks through the steps. Google's Policy Intelligence includes simulators for deny policies and Principal Access Boundary policies for this purpose.

Native boundaries versus a new layer

Boundaries can be enforced by the cloud itself or by a product in the request path. This index scores "Uses the cloud's own controls" at 10 percent, on the view that enforcement through mechanisms the customer already owns adds nothing new to fail. Sonrai says its enforcement uses the cloud's own organization-level policies with no proxy or broker, and Act says it deploys through pull requests to the customer's Terraform or CloudFormation using the cloud-native controls the customer already owns.

Whichever route a tool takes, the audit question is the same: where is the boundary written down, who can change it, and how would you know that someone did. A boundary stored only in a console is a boundary with no history. Version control is the simplest answer, since a pull request records the author, the reviewer and the date.

Sources

  1. AWS, permissions boundaries for IAM entities, read 10 Nov 2026
  2. Google Cloud, Principal Access Boundary policies, read 10 Nov 2026
  3. Microsoft Learn, Azure deny assignments, read 10 Nov 2026
  4. Sonrai Security, least privilege use case, read 10 Nov 2026
  5. Act Security, platform page, read 10 Nov 2026

Last reviewed 10 Nov 2026.