Azure / Entra Cloud Pentesting
Purpose
Full Azure + Entra ID offensive methodology: from a domain to a tenant model (black-box), or from a validated token/service principal (credentialed) to mapped identity, RBAC, resources, and escalation paths. Microsoft Graph and Azure CLI/API first — Sentinel/Defender are irrelevant to offensive workflow.
Entry Conditions
- Black-box: a target domain,
azurewebsites.net/blob.core.windows.netendpoints, or app client IDs fromcloud-attack-surface-discovery - Credentialed: ARM/Graph token, service-principal secret, or managed-identity token from
cloud-credential-abuse(audience already decoded)
Discovery Signals
- Tenant ID:
https://login.microsoftonline.com/<domain>/v2.0/.well-known/openid-configuration(returns tenant ID for the domain) - App/client IDs in JavaScript (MSAL configs,
clientId:), redirect URIs revealing app registrations - Azure-hosted hostnames (App Service, Functions, APIM, Front Door)
- Subscription/resource-group names in error messages, ARM URLs
Attack Model
tenant discovered → identity model mapped (users/SPs/apps/MI)
→ Graph permissions + RBAC roles enumerated
→ resource access (Key Vault, Storage, VMs, Automation)
→ escalation candidates → verified impact
Methodology
Step 1: Tenant Discovery (black-box)
curl -sk "https://login.microsoftonline.com/<target-domain>/v2.0/.well-known/openid-configuration" | jq ".token_endpoint, .issuer"
# issuer reveals the tenant ID; .well-known endpoints enumerate federation metadata
# Brute/verify common subdomains: <name>.azurewebsites.net, <name>.blob.core.windows.net
# MSAL JS configs in bundles: tenant + clientId + scopes → the app's consent surface
Step 2: Identity Model (Graph)
# With a Graph-audience token:
az rest --url "https://graph.microsoft.com/v1.0/me"
az rest --url "https://graph.microsoft.com/v1.0/users?$count=true" -H "ConsistencyLevel: eventual" # if Directory.Read*
az rest --url "https://graph.microsoft.com/v1.0/servicePrincipals" # enterprise apps
az rest --url "https://graph.microsoft.com/v1.0/applications" # app registrations
az rest --url "https://graph.microsoft.com/v1.0/oauth2PermissionGrants" # consent grants (delegated)
az rest --url "https://graph.microsoft.com/v1.0/policies/authorizationPolicy" # default user roles
# Decode your own token: roles claim (app roles) vs scp (delegated scopes) — these are DIFFERENT privilege models
Step 3: ARM / RBAC
# With an ARM-audience token:
az login --service-principal -u <appId> -p <secret> --tenant <tid> # or use the token
az account list
az role assignment list --all --output table # where does this principal have roles?
# Owner/Contributor/User Access Administrator on a scope = the RBAC escalation primitives
Step 4: Resource Access
# Key Vault (its OWN audience — ARM role ≠ data plane, but `Access Policies`/RBAC model decides):
az keyvault list && az keyvault secret list --vault-name <v>
# Storage: az storage container list --account-name <acct> (keys or SAS)
# App Service: az webapp config appsettings list (connection strings galore)
# VMs + Run Command (the classic exec primitive):
az vm run-command invoke -g <rg> -n <vm> --command-id RunShellScript --scripts "id"
# Automation accounts: runbooks + credentials
# Managed identities: az identity list; each maps to an ARM principal
Step 5: Chains
- Keys/SAS in Key Vault or app settings →
cloud-secrets-data-accessand resource access - VM Run Command / Automation → workload execution →
cloud-metadata-workload-identitylogic inverted (you ARE the workload) →azure-privilege-escalation - Cross-tenant apps/consent →
cloud-cross-account-tenant-trust
Token/Audience Rule
Before declaring a credential weak or powerful: determine aud (Graph / ARM / Key Vault / Storage), tenant, principal (user vs service principal vs managed identity), scopes (scp, delegated) vs roles (roles, application). A Graph token does not grant ARM access; an ARM token does not grant Graph privileges; Key Vault and Storage have separate audiences. Mint the right audience from the same credential when the identity supports it.
Evidence Contract
candidate (identity/permission/resource)
→ token with the right audience proves live access
→ concrete data or action (secret listed/read, VM command output, role assignment enumerated)
→ report with tenant/principal context
NOT evidence: tenant ID known, app registration existing, decoded token claims without a live call, az account list failing due to audience confusion (try Graph next)
Common Misses
- Audience confusion — Graph tokens tested against ARM and vice versa
- Consent grants (delegated) never enumerated — an app a user consented to is an implicit lateral path
- Managed identities under-used — their tokens mint per-resource audiences
- App settings/connection strings on App Service never listed
- Run Command forgotten as the exec primitive (no VM login required)
- Federated identity credentials on apps never inspected (token minting as the app)
False Positives / Non-Findings
- Multi-tenant apps visible in your tenant with no consent/grants (informational)
- Tokens expired or restricted by CA policies
- Service principals with no role assignments anywhere (report as unused credential exposure only)
Xalgorix Tool Strategy
terminal_executewith az CLI +az rest(Graph) — the baseline; no proprietary tools required- Python for JWT decoding and MSAL token minting per audience
- Ledger: tenant, principals, token audiences, role assignments, verified resource accesses
Specialist Handoffs
- Escalation →
azure-privilege-escalation - Kubernetes (AKS) → Entra integration here, RBAC/internals →
container-security - Cross-tenant trust →
cloud-cross-account-tenant-trust - Secrets →
cloud-secrets-data-access
Stopping Rule
Stop when: the tenant's identity model (users/SPs/apps/MIs), your effective Graph scopes and ARM roles, and the target's high-value resources (Key Vault, Storage, VMs, Automation) are mapped, and the strongest escalation candidates are verified or constraint-blocked.
