Most organisations use far more SaaS applications than their IT and security teams know about, and every unknown app is an unmanaged identity store holding company data and, often, an OAuth grant into core systems. Shadow SaaS is any cloud application used for work without going through procurement, security review or the identity provider. It arrives through free trials, personal sign-ups with a work email, "Sign in with Google" or "Sign in with Microsoft" buttons, browser extensions and, increasingly, AI tools. This guide explains the discovery sources that actually find these apps (identity provider logs, OAuth grants, network and endpoint telemetry, email and finance data, the browser), how to combine them into one inventory, how to triage and decide, and how to measure progress. It is for security engineers, IAM teams and SaaS owners. For the wider SaaS control set see the SaaS Security Guide.
Key takeaways
- No single source sees everything. Identity provider logs see federated and OAuth-connected apps, network logs see traffic, email sees sign-ups, finance sees paid tools and the browser sees what users actually do. Combine at least three.
- OAuth grants are the highest-risk shadow SaaS. An app a user connected to the mail or file platform holds a token into company data that survives the user's password change.
- Discovery is a lifecycle, not a project. New apps appear every week; run discovery continuously and feed findings into review, sanctioning or blocking.
- Restrict consent before you chase the backlog. Limiting user consent to verified publishers and low-impact permissions stops the inventory growing while you work through it.
- Standards are coming but not here yet. The OpenID Foundation IPSIE working group lists application discovery as possible future scope; today discovery relies on vendor features and your own correlation.
What counts as shadow SaaS
| Type | Example | Identity footprint | Typical risk |
|---|---|---|---|
| Local-account sign-up | User registers with work email and a password | A password account outside SSO; survives offboarding | Data left in an unmanaged tenant; reused passwords |
| Social or SSO sign-in to an unsanctioned app | "Sign in with Google" on a new tool | OIDC sign-in, sometimes with extra OAuth scopes | App visible in IdP logs but not reviewed |
| OAuth-connected app | A scheduling tool granted calendar and mail access | Delegated OAuth grant and refresh token | Direct, persistent access to company data |
| Browser extension | Writing assistant reading every page | Extension permissions plus the user's sessions | Page content and session theft |
| Shadow AI tool | Consumer AI assistant with document upload | Any of the above | Sensitive data in prompts and training data |
| Unmanaged tenant of a sanctioned app | A team's own instance of a file-sharing service | Separate admin and user accounts | Uncontrolled sharing settings and retention |
The Vercel incident in 2026 shows the OAuth case clearly: an employee connected a consumer AI office suite to a corporate Google account with broad access, attackers later stole that vendor's OAuth tokens, and one of them gave access to the employee's Workspace account. See the Vercel Context.ai OAuth Breach. For the AI-specific variant see the Shadow AI Discovery Guide.
Discovery sources compared
| Source | What it finds | What it misses | Effort |
|---|---|---|---|
| IdP sign-in logs | Apps using SSO or social sign-in through the corporate IdP, with user counts | Local-account sign-ups, personal accounts | Low: data already exists |
| OAuth grant inventory (Entra ID, Google Workspace) | Third-party apps with delegated or application permissions and their scopes | Apps that never touch the platform's APIs | Low to medium |
| Network logs (firewall, proxy, secure web gateway, DNS) | Traffic to known SaaS domains, volumes, users or IPs | Off-network devices without an agent; apps on shared CDN domains | Medium: parsing and catalogue matching |
| Endpoint agent telemetry | Same as network, but off-network and per device | Unmanaged and personal devices | Medium |
| Email signals | Welcome, verification, invoice and invitation emails from SaaS vendors | Apps that send no email | Medium: needs mail search and privacy review |
| Finance and expense data | Paid subscriptions on corporate cards and expenses | Free tiers | Low, needs finance cooperation |
| Browser management or browser security tooling | Installed extensions, sign-ins and in-browser use | Unmanaged browsers | Medium to high |
| User survey and attestation | Context: owner, purpose, data types | Anything users forget or hide | Low, but incomplete |
Identity provider logs
- Microsoft Entra ID: the sign-in logs include interactive, non-interactive, service principal and managed identity sign-ins, each recording who signed in, which client application was used and which resource was accessed. Group by application to find apps that are not in your approved list.
- Google Workspace: the Admin console under Security, Access and data control, API controls lists configured apps, accessed apps (apps that have used Google data) and apps pending review, with the number of users, the requested services and the OAuth scopes. Lists can be exported as CSV, and new apps typically appear 24 to 48 hours after authorisation.
- Other IdPs: any IdP that logs application assignment and sign-in events can do the same; export them to your SIEM so they can be joined with other sources.
OAuth grant inventory
In Entra ID, the Permissions blade of each enterprise application shows admin-consented and user-consented grants. At scale, query Microsoft Graph: oauth2PermissionGrants returns delegated grants and appRoleAssignments returns application permissions. Sort by scope sensitivity (mail, files, directory write, offline access) and by number of consenting users. In Google Workspace, the accessed apps list gives the same view with scopes per app. See the SaaS and OAuth App Governance Guide and the Consent Phishing and Illicit OAuth Grants Guide.
Network and endpoint telemetry
Cloud discovery features in CASB and secure web gateway products match traffic logs against an app catalogue. Microsoft Defender for Cloud Apps, for example, documents a catalogue of more than 31,000 apps scored on more than 90 risk factors, ingestion from more than 40 firewall, proxy and gateway formats through manual upload, a log collector (Syslog or FTP) or endpoint integration, and a distinction between one-off snapshot reports and continuous reports. Continuous, agent-based collection matters because remote staff are often off the corporate network. See the CASB Guide and the Secure Web Gateway and Web Filtering Guide.
Email, finance and the browser
- Email: searching inbound mail for vendor welcome, verification and invoice messages finds local-account sign-ups that no identity log can see. Agree the scope with privacy and HR first: search sender domains and subject patterns, not message bodies.
- Finance: expense and card feeds reveal paid tools and their business owners, which is useful for triage.
- Browser: managed browsers report installed extensions, and policies such as Chrome's
ExtensionSettingscan block extensions by ID or by the permissions they request. See the Browser Extension Security Guide.
Building one inventory
- Normalise the app identity: map domains, OAuth client IDs, IdP application IDs and vendor names to one canonical app record. A single vendor often has several domains and several client IDs.
- Attach users: for each app, record which users appear in which source and when last seen.
- Attach access: OAuth scopes, whether the grant is delegated or application-level, whether offline access (refresh tokens) is held.
- Attach context: business owner, data types, vendor risk rating, contract status.
- Feed the CMDB or SaaS register: the output becomes the software and services inventory that NIST CSF 2.0 ID.AM-02 and ID.AM-04 expect. See the Asset Inventory and CMDB Security Guide.
Triage and decision guidance
| Finding | Default decision | Action |
|---|---|---|
| Many users, low-risk scopes, legitimate business need | Sanction | Put it behind SSO, assign an owner, add SCIM where available, move to a managed tenant |
| Duplicates an approved tool | Migrate and retire | Tell users the approved alternative, set a cut-off date, then block |
| High-risk scopes (mail read, full drive, directory write) from an unverified publisher | Revoke and block | Revoke grants, block the client ID, contact the users |
| One or two users, no data access | Monitor | Leave unconfigured but visible; review if user count grows |
| Consumer AI tool handling company data | Decide under the AI use policy | Approve an enterprise tier or block; see the Shadow AI Policy Template |
| Unknown app with application-level permissions | Treat as an incident until explained | Investigate who consented and what was accessed |
- Small organisations: start with the IdP and OAuth grant views you already own, plus the card statement. That covers the riskiest apps for no extra spend.
- Large organisations: add continuous network or endpoint discovery and browser telemetry, and give each business unit a SaaS owner who approves or rejects findings.
- Blocking versus enabling: blocking without an alternative drives use to personal devices, where you see nothing. Publish an approved catalogue and a fast request route alongside any block.
Preventive controls that stop the growth
- Restrict user consent: in Entra ID, Microsoft recommends allowing user consent only for apps from verified publishers and only for permissions classified as low impact (the
microsoft-user-default-lowpolicy), with the admin consent workflow enabled so users can request more. Changes apply to future consent only: existing grants must be reviewed and revoked separately, and a revoked grant can be re-consented unless consent is blocked. - Google Workspace access levels: set apps to Trusted, Limited, Specific Google data or Blocked, decide how unconfigured third-party apps are treated (allow, allow only basic sign-in, or block), and mark sensitive services such as Gmail and Drive as Restricted, which stops untrusted apps and revokes their tokens.
- Browser policy: block extensions that request high-risk permissions and allow-list the ones the business needs. The Cyberhaven Chrome Extension Breach shows how a trusted extension can turn hostile after a single OAuth consent by its publisher.
- Offboarding: include OAuth grants and known shadow apps in leaver processes. See the Employee Offboarding Security Guide.
Where identity fits, and where standards are going
Every shadow SaaS app is an identity problem: a local account outside SSO, a user's OAuth grant acting as a long-lived non-human credential, or an integration token held by a vendor. Treat the refresh tokens held by third-party apps as NHIs with owners, scopes and expiry. See the API Keys vs OAuth Tokens Guide.
The OpenID Foundation's IPSIE (Interoperability Profiling for Secure Identity in the Enterprise) working group profiles OpenID Connect, OAuth, SCIM and Shared Signals so that SaaS apps and identity providers interoperate with secure defaults. Its initial scope is single sign-on, lifecycle management, entitlements, risk signal sharing, logout and token revocation; its charter page lists discovering applications used within an enterprise, and monitoring and managing application usage, as topics it may also address. Nothing in IPSIE gives you discovery today, but apps that adopt IPSIE profiles will be easier to bring under SSO, provisioning and revocation once found. See the IPSIE (Interoperability Profiling for Secure Identity in the Enterprise) Guide.
Implementation roadmap
- Weeks 1 to 2, baseline: export IdP application lists and OAuth grants from Entra ID and Google Workspace; pull a year of card and expense data. Produce the first inventory.
- Weeks 2 to 4, stop the growth: restrict user consent, enable admin consent requests, set unconfigured app handling in Google Workspace. Pilot on one business unit first and watch the request volume for a week before going tenant-wide. Roll back by restoring the previous consent policy if critical workflows break.
- Weeks 4 to 8, revoke the worst: revoke high-risk grants from unverified publishers, after notifying users.
- Months 2 to 4, add telemetry: continuous network or endpoint discovery, browser extension reporting, email signal search with privacy approval.
- Ongoing: weekly triage of new apps, quarterly attestation by SaaS owners, annual review of the approved catalogue.
Detection and measurement
- Alert on: new consent to mail, file or directory-write scopes; application-permission grants; an app seen for the first time with more than a threshold of users in a day; large uploads to unsanctioned cloud storage (ATT&CK T1567.002); new browser extensions with broad host permissions.
- Investigate token theft paths: stolen application access tokens map to ATT&CK T1528 Steal Application Access Token.
| Metric | Why it matters |
|---|---|
| Apps in inventory with a named owner (%) | Ownership is what makes review possible |
| OAuth grants with high-risk scopes from unverified publishers | Should trend to zero |
| Median days from first sighting to decision | Shows the triage process keeps up |
| Share of sanctioned apps behind SSO | Measures the conversion from shadow to managed |
| Discovery source coverage (sources feeding the inventory) | Fewer than three leaves known blind spots |
| Leavers with remaining third-party grants after 7 days | Tests offboarding of shadow access |
Framework mapping
| Framework | Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | ID.AM-02, ID.AM-04, GV.SC-04, DE.CM-03 | Software and supplier service inventory, supplier criticality, monitoring use |
| NIST SP 800-53 Rev. 5 | CM-8, CM-10, CM-11, SA-9, AC-20 | Inventory, software usage restrictions, user-installed software, external services, use of external systems |
| CIS Controls v8.1 | Control 2 Inventory and Control of Software Assets | Only authorised software in use; see the CIS Control 2 Inventory and Control of Software Assets Guide |
| MITRE ATT&CK | T1528, T1567.002, T1176.001 | Token theft, exfiltration to cloud storage, malicious browser extensions |
Common failure patterns
- Network-only discovery: relying on proxy logs while most staff work remotely, or while the riskiest apps are OAuth-connected and talk server to server, never through your network.
- Inventory without decisions: a dashboard listing hundreds of apps that nobody owns or triages.
- Revoking without restricting: revoked grants come back because user consent is still open.
- Ignoring tokens held by vendors: the risk is not only what the app does today but what an attacker can do with its stored tokens after breaching the vendor, as the Vercel Context.ai OAuth Breach showed.
- Blocking with no alternative: users move to personal accounts and devices, and the data goes with them.
- Email search without governance: mailbox scanning started without privacy and works council review, which halts the programme.
Practitioner checklist
- Export application and OAuth grant lists from every IdP and collaboration platform monthly.
- Restrict user consent to verified publishers and low-impact permissions, with an admin consent request route.
- Set a default for unconfigured third-party apps in Google Workspace and restrict Gmail and Drive.
- Feed at least three discovery sources into one normalised inventory with owners.
- Triage new apps weekly against a published decision table.
- Revoke high-risk grants and block their client IDs; confirm they do not return.
- Manage browser extensions by allow-list or by blocked permissions.
- Add OAuth grants and known shadow apps to leaver processes.
- Report the six metrics above to the SaaS governance forum each quarter.
Standards and references
- OpenID Foundation: IPSIE Working Group
- NIST: Cybersecurity Framework 2.0
- CIS: Control 2 Inventory and Control of Software Assets
- Microsoft Learn: Configure how users consent to applications
- Microsoft Learn: Review permissions granted to enterprise applications
- Google Workspace Admin Help: Control which third-party and internal apps access Google Workspace data
- Microsoft Learn: Set up Cloud Discovery in Defender for Cloud Apps
- Google Chrome Enterprise Help: Manage apps and extensions with Group Policy
Related NHI Mgmt Group resources: SaaS Security Guide · SaaS and OAuth App Governance Guide · Shadow AI Discovery Guide · Top 10 SaaS Security Risks · CASB Guide