Skip to content
Get started

How Permafrost works

Permafrost models Azure RBAC and Entra permissions separately. Each has its own assignments, evidence and remediation paths. Collection and analysis are read-only, and your data is scoped to your customer.

Two permission surfaces in Azure

The common CIEM-category mistake is to treat Azure as one permission surface. It is not. Microsoft Cloud has two. They run on different consent flows, surface in different log streams, and live behind different APIs.

The first surface is Azure RBAC: the control plane where role assignments at management group, subscription, resource group, and resource scope decide who can do what to the customer’s Azure resources. Azure RBAC assignments are visible in the Azure Resource Manager APIs and audited in the ARM activity log.

The second surface is Microsoft Graph consent: the directory plane where application permissions are granted to enterprise apps and service principals. Graph consents are visible in the Microsoft Graph APIs and audited in the Entra directory audit log.

Permafrost treats them as distinct first-class surfaces. The semantics, the blast radius, and the remediation pattern on each side are different enough that collapsing them produces worse answers, not simpler ones.

Why this matters

Azure RBAC permission-gap analysis compares collected assignments with observed control-plane activity. You can inspect the role, its scope and the operations observed during the analysis window.

Activity-based findings cite collected Azure activity-log evidence; configuration findings cite collected configuration. Microsoft Graph application permissions and Entra directory roles have their own findings, preserving the context of each permission surface.

Unused Permission % is one factor inside Unified Principal Risk (UPR), the 0–100 score Permafrost computes for every security principal. UPR combines privilege tier, unused standing permission, blast-radius scope, dormancy, sensitive permissions, PIM and Conditional Access posture, and that principal’s open findings. Disabled-account state and active privileged sign-in risk also contribute. The breakdown shows the contribution of each factor.

The data flow, conceptually

Three steps. Read-only throughout. Each step is one-way: Permafrost reaches into the customer’s data plane through Reader-class scopes and nothing more. The customer’s tenant is never modified by the ingest or analyze passes.

Permafrost collects identity and entitlement data with read-only access. Workload identity federation uses no stored customer credentials; the optional client-secret connection stores its secret in Vault. Supported Mode C actions require a delegated token authorized by you. It travels through your browser in a sealed session only the server can open. Permafrost accepts it once within ten minutes and stores no credential.

  1. Ingest. Permafrost pulls identity, permission, and activity data from Microsoft Graph and Azure Resource Manager. The pull cadence depends on the data type and the customer’s tier; identity inventory refreshes more often than slow-changing role-policy data.
  2. Analyze. The ingested data is partitioned by customer (more on this below) and fed into three analysis families: permission-gap analysis (the engine behind the unused-permission factor of UPR), deterministic finding detection (privilege paths, toxic combinations, dormant high-privilege accounts), and pattern matching that powers the incident timeline.
  3. Surface. The analysis output is presented through dashboards (Identities, Findings, Roles, PIM, Activity), exportable artifacts (custom Azure roles as ARM / Bicep / Terraform), and the remediation options supported by each finding: manual guidance, downloadable scripts and delegated actions.

What gets ingested

The analysis runs on three data families.

  • Identity inventory. Users (members and guests), security groups, service principals (application identities), managed identities (system-assigned and user-assigned), and AI agent identities (the copilot-class principals that have appeared in modern Azure tenants).
  • Permission inventory. Azure RBAC role assignments at every scope, Microsoft Graph application permissions, Privileged Identity Management eligible assignments and their activation policies, and Conditional Access policy context that affects the access picture.
  • Activity inventory. ARM activity-log entries for control-plane operations on Azure resources, and the Entra directory audit log for directory-side changes. We pull enough to evidence permission-gap findings and the incident-timeline patterns; we are not a log retention destination.

Per-customer tenant isolation

Permafrost supports customers connecting more than one Azure tenant. Each customer organization owns a workspace. Inside that workspace, the customer connects one or more Azure tenants through read-only consent. The dashboard rolls up across those connected tenants; each tenant is analyzed independently underneath.

The isolation rule has two parts.

Within a customer’s workspace, Permafrost can correlate findings across the tenants that customer has connected. A privileged identity that appears in two of the customer’s connected tenants surfaces in both. The dashboard view rolls up across the tenants that customer has authorized.

Across separate customer organizations, Permafrost does not aggregate. There is no cross-workspace comparison, no benchmark feature that exposes one customer’s data to another, no shared join key that can span workspaces. Every dataset is partitioned by customer at the storage layer; the partition is structural, not just a query-time filter.

When this documentation says “your connected tenants” it means the set of tenants this customer has connected to this customer’s workspace. It never means anything broader than that.

Coverage

The Azure activity collector requests a 90-day initial window, then resumes from the latest stored event with a two-hour overlap. Activity-based analysis describes the operations observed in that window, not complete data-plane usage or future workload needs.

When effective-permission data is absent, the analyzer uses an identity-type heuristic for the unused-permission input. It is not a measured usage percentage. Use the source evidence and workload requirements together when planning an access reduction.