One graph for every identity, and everything it can reach.
Human, non-human and AI identities across your cloud: the access you granted, the reach they really have, and the evidence behind every answer.
Early access · Read-only access · Plans from $9,000/year
The decision surface. One finding, its band, the one move — and the basis every number was derived from.
Access is what you granted. Reach is what they get.
Every human, non-human, and AI identity in your cloud reaches further than the role you assigned it. AuditGraph computes the difference — from your architecture, before an attacker does. Pick a role, then widen the scope it is assigned at.
Reader
Reads the configuration of everything in scope. Cannot change anything, and cannot open the data — but it maps where the data is.
Contributor
Creates and changes everything in scope except access. That includes listing storage account keys, which opens the data.
Owner
Everything Contributor can do, plus granting access. An identity with Owner can give any identity — including a new one — Owner.
Azure role assignments inherit downward. A role granted at a management group applies to every subscription, resource group and resource beneath it — today and every one created later. The inventory lists one assignment; the reach is everything under its scope.
Reader assigned at management group reaches
scope: /providers/Microsoft.Management/managementGroups/{mg}
Reader assigned at subscription reaches
scope: /subscriptions/{id}
Reader assigned at resource group reaches
scope: /subscriptions/{id}/resourceGroups/rg-app
Reader assigned at resource reaches
scope: /subscriptions/{id}/resourceGroups/rg-app/providers/Microsoft.Storage/storageAccounts/{name}
Contributor assigned at management group reaches
scope: /providers/Microsoft.Management/managementGroups/{mg}
Contributor assigned at subscription reaches
scope: /subscriptions/{id}
Contributor assigned at resource group reaches
scope: /subscriptions/{id}/resourceGroups/rg-app
Contributor assigned at resource reaches
scope: /subscriptions/{id}/resourceGroups/rg-app/providers/Microsoft.Storage/storageAccounts/{name}
Owner assigned at management group reaches
scope: /providers/Microsoft.Management/managementGroups/{mg}
Owner assigned at subscription reaches
scope: /subscriptions/{id}
Owner assigned at resource group reaches
scope: /subscriptions/{id}/resourceGroups/rg-app
Owner assigned at resource reaches
scope: /subscriptions/{id}/resourceGroups/rg-app/providers/Microsoft.Storage/storageAccounts/{name}
- rank 4 Management group 1,480 resources above the assignment
- rank 3 Subscription 312 resources in reach
- rank 2 Resource group 14 resources in reach
- rank 1 Resource 1 resource in reach
Sample estate · counts are illustrative. Scope ranks (resource 1 → management group 4) are the product's blast-radius policy.
One compromised credential. Six tiers of reach.
When a credential is compromised, the incident question is always the same: what is the chain from that credential to the data we lost? That chain crosses humans, non-humans, AI agents, models, and datasets. Most tools see one tier of it.
Every edge is a verified RBAC, OIDC trust, or data-classification relationship — read from your tenant's configuration. Patent provisional filed June 2026.
Accurate. Also the part you could already query yourself.
The full authority path — including the tiers that have no owner today.
One leaked secret. Now watch one path.
From a credential to the data at the end of its reach, one hop at a time. Each hop is a relationship read from configuration, with the evidence it rests on.
- StartLeaked app secret Credential
A client secret on an app registration, committed to a build log.
evidence · application credential · key id, created, expires · collected at publication #72
- Hop 1Service principal Non-human identity
The secret signs in as the app’s service principal. No owner is recorded.
evidence · service principal · app id · owner: none recorded · collected at publication #72
- Hop 2Owner at subscription Role assignment
The service principal holds Owner on the production subscription.
evidence · role assignment · scope /subscriptions/{id} · role definition Owner · collected at publication #72
- Hop 3Key vault Resource · inherited
Owner inherits to the payments key vault and can grant itself access to its secrets.
evidence · inherited from scope /subscriptions/{id} · Microsoft.KeyVault/vaults · collected at publication #72
- Hop 4Patient data store Classified data · PHI
The same inherited Owner reaches the storage account that holds patient records.
evidence · data trust zone PHI (declared) · inherited from scope /subscriptions/{id} · collected at publication #72
One move closes it: scope the service principal from subscription Owner to the one resource it needs, and rotate the secret.
Sample tenant. A candidate path is derived by rules from configuration: each hop is a relationship AuditGraph collected, not an observed attack. Nothing here is a customer finding.
You have an inventory. An attacker has a path.
The market stops at truth: a list of principals, roles, permissions, misconfigurations, graph edges. All accurate — and all irrelevant to the person who has sixty seconds and a board meeting.
Truth and relevance are different properties. Detection is commoditized. The decision is the moat.
The tiers nobody owns
SailPoint stops at humans. CIEM stops at roles. NHI vendors stop at service accounts. AI security tools stop at runtime. Each covers one tier; none connects the chain.
Log-based tools miss static risk
An identity that has never signed in produces no telemetry — and can still hold Owner on a production subscription. Absence of activity is not absence of authority.
Non-human identities outnumber humans, and nobody is counting
Service principals, managed identities, workload and CI/CD identities, PATs, OAuth apps, AI agents. They accumulate permissions and rarely lose them.
Counts are not consequences
"You have 4,000 identities" is not a decision. "This one reaches regulated data and three people can assume it" is.
Reach is in your architecture. It was never in your logs.
Roughly 70% of organizations do not run the logging a behavior-based product depends on. AuditGraph derives risk from what is already there — role assignments, PIM, federated credentials, RBAC scopes, ARM, GitHub OIDC, network configuration.
Business-first
The headline is the business condition, never the role name. The role is how we prove it; the consequence is what we say.
Every number has provenance
Each value carries its basis — measured, computed, projected, estimated, attested, or unknown. A value without a basis does not render.
Unknown is a real answer
Absence is never painted as safety, never coerced to zero, never shown green. "We cannot see this" is a finding, not a gap to fill.
Deterministic, not generative
No language model sits in any reasoning, ranking, or narration path. Argus formats conclusions the engines reached. It cannot improvise.
A score you cannot decompose is a score you cannot defend.
A single 0–10 number tells a CISO nothing they can act on and nothing they can defend to an auditor. Worse, it hides its own arithmetic — and a score you cannot decompose is a score you cannot argue with.
AuditGraph retired numeric severity scoring entirely. In its place: bands derived from checkable facts — the privilege tier an identity actually holds, the reach that privilege actually grants, the governance controls that are actually absent. Every band decomposes back to the assignments that produced it.
When your auditor asks where a rating came from, the answer is a scope string and a role definition — not a weighting we chose.
Silence is not safety. Every identity gets a verdict anyway.
Static association analysis assigns each identity in your tenant one of nine verdicts — continuously, with no dependency on sign-in telemetry.
Non-human identities default to risky-until-proven. On an NHI, the absence of a risk signal is itself a signal — never a neutral gap.
One graph. Four ways to act on it.
The graph is the evidence, never the headline. Each surface is a different question asked of the same canonical facts.
Identity Graph
Relationships and reachability across the whole estate. Who can assume what, through which trust, to reach which resource. The depth that makes every other answer credible.
Explore the graph →Data reachability
The resource-first lens. Start at a storage account, a database, a dataset — and see every identity that can reach it, by what path, at what privilege.
See data exposure →AI Identities
AI agents are a subtype of non-human identity, not a separate product. The same graph, the same verdicts, extended to agents, models, and the data they can reach.
AI identity coverage →Attack paths & exposure
Blast radius and data reachability computed before exploitation, from configuration rather than from an incident. Ranked by business consequence.
Full platform →Four lenses, one graph. Your question first.
The Executive Summary opens on the question your role asks — and answers it in one sentence and three numbers. Switch lens and the question changes; the facts do not.
What is putting the business at risk, and what must leadership decide?
Identity exposure got worse since the last snapshot. The decision: move standing Tier-0 and Tier-1 roles to just-in-time (PIM-eligible).
Where is privilege too broad, and which design change removes it?
Authority concentrates in standing assignments at the broadest scopes. The design change that removes the most: just-in-time for Tier-0 and Tier-1.
What changed since the last scan, and what do we work first?
Three identities gained admin rights since the previous snapshot and two new critical findings appeared. Confirm the grants first.
Are we audit-ready?
Most mapped controls have evidence, and most of those still show a gap to close before an audit. The latest evidence pack is sealed and checkable.
All four read the same published snapshot. Illustrative numbers from the product’s demo tenant.
Take the product tour — seven real screens on demo data.
Azure is generally available today. AWS and GCP are roadmap.
We do not claim coverage our connectors do not have. Here is exactly where the product stands today.
Validated to a documented 100,000-identity / 950,000-role-assignment baseline. Larger estates should be sized with us before onboarding.
- Microsoft Entra ID
- Microsoft Azure
Read-only. Azure is the generally available cloud; AWS and GCP are roadmap.
- Slack
- Microsoft Teams
- Signed webhooks
Change alerts on every plan, including the evaluation.
- Jira
- ServiceNow
- Azure DevOps
Ticket workflows on Professional and Enterprise.
- Azure Monitor
- Microsoft Sentinel
- Splunk
- Datadog
- Elastic
- Sumo Logic
Findings export on Professional and Enterprise.
Product names are trademarks of their owners and are used only to describe compatibility. No endorsement or partnership is implied.
What CISOs ask before they approve.
Does AuditGraph change anything in our tenant?
No, not on its own, and that is a permanent product decision. The only change is one your administrator makes at sign-up: granting AuditGraph's collector the read-only Azure Reader role, once, at the tenant root management group, so it covers every current and future subscription. The administrator can let AuditGraph make that one assignment with their own one-time sign-in, or run the copy-ready Azure CLI, PowerShell or portal steps themselves, or assign Reader only to selected subscriptions or management groups. You can remove the role at any time in Access control (IAM) at the tenant root management group. Fixes are always applied by your own team, with the steps and scripts AuditGraph provides.
What exactly are we approving on the Microsoft consent screen?
Five read-only Microsoft Graph application permissions for the AuditGraph collector. Directory.Read.All: users, groups, applications, service principals, directory roles and the app consents your users have granted. Application.Read.All: app registrations, enterprise applications and their permissions. AuditLog.Read.All: sign-in activity and MFA registration (method types only, never phone numbers), which Microsoft returns only with Entra ID P1 or P2. RoleManagement.Read.All: Entra role definitions, assignments and Privileged Identity Management eligibility. Policy.Read.All: Conditional Access policies and security defaults; AuditGraph never creates, changes or deletes a policy. Every permission is read-only, and nothing secret is typed into AuditGraph: no password, key or client secret.
Who in our organization must sign up?
An administrator of your Microsoft Entra tenant who holds Global Administrator or Privileged Role Administrator. The role must be active when they sign in: if it is only eligible through Privileged Identity Management, they activate it first. Personal Microsoft accounts are refused. Anyone else is asked to involve an administrator and gets a link they can share with them.
What does the free evaluation include, and what does it cost?
Evaluations open soon. Each one will be free for 14 days, with no credit card. It is read-only: you get the full identity graph, findings, attack paths and evidence for one Entra tenant, but workflow changes such as triage, risk acceptance and access reviews are not available. It includes 5 seats, the first scan and up to 3 rescans a day; there are no scheduled scans. Users sign in with their Microsoft work account; connecting your own identity provider for single sign-on needs a paid plan. Every customer will start here, and from day 10 the app shows an evaluation summary of what was found and what your team fixed.
Where is our data stored, and can we choose?
You will choose where your organization is created on the sign-up page, and each choice shows its residency statement before you confirm. Only regions open for sign-up are offered: the first is Microsoft Azure, Central US, and further regions open as they launch. AuditGraph stores no credential for your tenant. Questions you ask the Argus AI assistant are processed by Azure OpenAI global deployments, which may process prompts in any Azure region.
Is our data used to train AI models?
No. Your data is never used to train AI models. The Argus assistant runs on Azure OpenAI in our own Azure subscription, which does not use customer data to train models.
More answers for CISO, GRC, SOC and architecture teams in the FAQ.
Shape what we build next. From the inside.
We're working with a small group of early customers. Design partners get early access, direct input into the roadmap and founding pricing.
Find the reach you never granted.
A scoped assessment against your own Azure tenant. Read-only, always: AuditGraph never changes your tenant on its own.
Early access · Read-only access · Plans from $9,000/year
Read-only access · No agents · No log ingestion · Azure generally available