SaaS security checklist: test tenant isolation and keep the evidence buyers ask for
For the founder, CTO or technical lead of a software business that sells subscriptions to other businesses.
This guide follows a customer tenant from sign-up to cancellation, shows where one tenant's data can reach another, and gives you a two-tenant evidence grid to run on a staging copy of your own product, with fields for the scope, exceptions, remediation and retest status a buyer or auditor may ask to see.
On this page
Two-tenant isolation evidence grid
Run it on a staging copy of your own product, never on production and never on a system you do not own. Create two invented tenants with invented users: in Tenant A, a-owner, a-staff, a-leaver (a second staff member, used only for demotion and removal) and, if the role exists, a-billing; in Tenant B, b-owner and b-staff; and support-1 in the role your own staff use.
How to use itHide the instructions
Seed each tenant with invented customer records, files and an export whose names mark the owner, such as A-CUSTOMER-01 and B-CUSTOMER-01. Row G1 creates two more test tenants, Tenant C and Tenant C-West, for the cancellation row. Every sign-in uses a mailbox you control, and any email or text a row triggers goes to an address or number you control. Write the expected result for each row from your own rules before sending anything. Rows G1 to G13 and G15 sign in or send requests and carry their limits on the row; G14 only reads a policy and asks questions, and so does the storage question in G10. Where a row touches a third-party service, read that vendor's terms first and ask the vendor if they are unclear: the grid tests your product's side only. The last column holds four buyer-audit fields: scope (the routes, accounts, build and date), exceptions (any result that differed from the expected one, and any part not run, with the reason), remediation (the fix, who made it and the build it shipped in) and retest status (fixed, partially fixed, still reproducible or not retested). The grid has 15 rows, G1 to G15, and 8 columns.
15 rows
Step
Data involved
G1. Tenant provisioning through self-serve sign-up
- Data involved
- The new tenant record, its identifier, plan and owner account, and the details the owner enters.
- Who should reach it, and who should not
- The person signing up becomes owner of a new, empty tenant. Nobody gains a role in Tenant A or Tenant B by signing up.
- Test case and its limits
- Sign up Tenant C and Tenant C-West with mailboxes you control. Repeat one sign-up with a tenant id, role, plan or owner field added to the request, naming Tenant A. Limits: staging only, no more than four sign-ups. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Each sign-up creates a tenant whose identifier the server assigned. Added fields are ignored or refused, and the new owner sees no Tenant A records.
- Evidence to keep
- The sign-up requests and responses, the new tenant identifiers, and the new owner's empty record list.
- What this row cannot show
- Tenants your staff create through an admin screen or an import, and tenants created through a customer's own single sign-on. Each needs a row of its own.
- Buyer-audit fields
- Scope: the sign-up route and the tenant-creation step. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G2. Invitation sent
- Data involved
- The invitee's email address, the role offered, the tenant the invitation names and the invitation token.
- Who should reach it, and who should not
- An owner, or a role your rules allow, invites people into their own tenant at a role no higher than their own. Nobody invites people into another tenant.
- Test case and its limits
- As a-staff, send an invitation offering the owner role. As a-owner, send one with the tenant field changed to Tenant B. Invitations go only to mailboxes you control. Limits: staging only, no more than four invitations. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Both are refused, or the tenant and role come from the session and your rules rather than the request. WSTG-IDNT-03 asks whether an administrator or other user can provision accounts with privileges greater than their own.
- Evidence to keep
- Each request and response, and the invitation list for both tenants afterwards.
- What this row cannot show
- Bulk invitations, directory sync and invitations your staff create, unless each is added as a row.
- Buyer-audit fields
- Scope: the invitation routes, as a-owner and a-staff. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G3. Invitation accepted
- Data involved
- The invitation link and the account it attaches to the tenant.
- Who should reach it, and who should not
- The invited person, once, before the invitation expires or is withdrawn. Whether the accepting account must use the invited address is your rule: write it down first.
- Test case and its limits
- a-owner invites a mailbox you control. Open the link signed in as b-staff, then accept it as the invited address, then open it again. Withdraw a second invitation and open its link. Limits: staging only, two invitations, no more than six requests. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- b-staff is refused, or must sign in as the invited address, as your rule says. A used or withdrawn link no longer works, and nobody joins Tenant A without an invitation.
- Evidence to keep
- The response at each step, and Tenant A's member list after each one.
- What this row cannot show
- How long links stay valid in production, if staging uses other settings. Whether link tokens can be guessed needs code access or your developer's written answer.
- Buyer-audit fields
- Scope: the invitation acceptance route, as a-owner and b-staff. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G4. Owner and staff roles, including removal
- Data involved
- Tenant settings, the member list, role assignments and each member's open sessions.
- Who should reach it, and who should not
- The owner changes roles and removes members. Staff cannot raise their own role, remove the owner or reach Tenant B's settings.
- Test case and its limits
- As a-staff, send role-change requests for a-staff and for a-owner. As a-owner, demote a-leaver while a-leaver has a session open, then remove a-leaver, sending one request from a-leaver's session after each change. Limits: staging only, test accounts only, no more than eight requests. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- a-staff's own changes are refused. The demotion applies to a-leaver's next request, and after removal a-leaver reaches nothing in Tenant A. ASVS 5.0 requirement 8.3.2, at Level 3, asks for changes to the values behind authorisation decisions to apply immediately, and where they cannot, for mitigating controls that alert when a consumer acts after losing authorisation and revert the change.
- Evidence to keep
- Each request and response, with the time of each change and of the next request.
- What this row cannot show
- Whether a removed member's sessions end in other tenants they belong to, or in a sign-in service you do not run.
- Buyer-audit fields
- Scope: the role and membership routes, as a-owner, a-staff and a-leaver. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G5. Billing-only user
- Data involved
- The plan, the invoices, the billing contact and any payment details your product stores. Where a payment provider holds card details, they sit outside the grid.
- Who should reach it, and who should not
- a-billing reaches Tenant A's billing pages. It does not reach customer records, exports, member management or Tenant B. Staff reach billing only if your rules say so.
- Test case and its limits
- As a-billing, request a Tenant A customer record by id, the record list, an export and the invitation route. As a-staff, request the invoice list and the plan-change route without completing a change. Limits: staging only, no more than eight requests. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Each request outside the role is refused by the server, not only hidden in the screens. WSTG-IDNT-01 asks the tester to review the granularity of the roles and the needs behind the permissions given.
- Evidence to keep
- The permission matrix entries for both roles, and each request and response.
- What this row cannot show
- The payment provider's own pages and platform. Read its terms before any request reaches it.
- Buyer-audit fields
- Scope: the billing routes, and the routes refused to a-billing. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G6. Support access and impersonation
- Data involved
- Whatever the impersonated user can see in Tenant A, and the audit record of the support session.
- Who should reach it, and who should not
- support-1, only in the way your product decides: who may start it, whether the tenant is told, what it may change and when it ends. No tenant user can start it.
- Test case and its limits
- As support-1, start an impersonation of a-staff, view one record, then end it. As a-owner and as b-owner, call the route that starts impersonation. Limits: staging only, test accounts only, never a real user's account, no more than six requests. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Both owners are refused. The audit record names support-1 as the actor as well as a-staff, the tenant and the time, and the impersonation ends when support-1 ends it.
- Evidence to keep
- The audit entries for the support session, and the refusals for both owners.
- What this row cannot show
- Support access from outside the product, such as direct database queries or a console your hosting provider supplies.
- Buyer-audit fields
- Scope: the impersonation start and end routes, as support-1, a-owner and b-owner. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G7. Exports and export jobs
- Data involved
- Export files of a tenant's records, and the background job that builds each file where exports run that way.
- Who should reach it, and who should not
- The roles your permission matrix allows, for their own tenant only.
- Test case and its limits
- As b-owner, request an export with any tenant or account parameter changed to Tenant A. As b-owner, request the status and the download of a finished Tenant A export job by its id. Count the marker names in each file. Limits: staging only, invented records only, no more than six requests. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Tenant B gets only B-marked rows, and Tenant A's job and file are refused. OWASP's Authorization Regression Testing Cheat Sheet treats even a single leaked record identifier as a critical failure.
- Evidence to keep
- Row counts by marker for each file, and the job responses. Keep files on the test machine only and delete them afterwards.
- What this row cannot show
- Which columns an export should hold, which the sensitive data exposure guide covers.
- Buyer-audit fields
- Scope: the export and export job routes, as b-owner. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G8. Files and object links
- Data involved
- Uploaded files, generated documents, and the links the product hands out for them.
- Who should reach it, and who should not
- Users of the owning tenant, for as long as your rules allow. If your rules say a link ends with its holder's access, it does.
- Test case and its limits
- Before this row, a-owner invites a-leaver again at the staff role, to a mailbox you control. Take a link a-leaver is given for an invented Tenant A file and request it signed in as b-staff, with no session, after a-owner removes a-leaver, and after the link's stated expiry. Change the file identifier in a valid link to another Tenant A file. Limits: staging only, invented files only, no more than ten requests. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Each request your rules refuse is refused, and a link opens only the one file it was issued for. The Multi-Tenant cheat sheet asks for signed URLs restricted to the required object and operation, with an appropriate expiration and revocation model.
- Evidence to keep
- The file each link was issued for, never the link itself, with the response status and the time of each request.
- What this row cannot show
- Whether a link's lifetime suits the data, which is your decision, and the storage provider's own access settings.
- Buyer-audit fields
- Scope: the file and document routes, as a-owner, a-leaver and b-staff. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G9. Search, suggestions, counts and dashboards
- Data involved
- Search results, type-ahead suggestions, record counts and dashboard figures, including any the product caches.
- Who should reach it, and who should not
- Each user sees results from their own tenant only.
- Test case and its limits
- As b-staff, search for A-CUSTOMER and type its first letters into every search box. Load the same dashboard as a-staff and then as b-staff, one straight after the other. Limits: staging only, no more than ten requests. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- No Tenant A name, count or figure reaches b-staff. The Multi-Tenant cheat sheet asks for the tenant identifier in every cache key whose value or authorisation varies by tenant.
- Evidence to keep
- The search responses and both dashboard screenshots.
- What this row cannot show
- A cache fault that appears only under production traffic, or only after a deployment empties the cache.
- Buyer-audit fields
- Scope: every search box and dashboard, as a-staff and b-staff. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G10. API tokens
- Data involved
- Tokens issued to a tenant or a user, and the routes each one can call.
- Who should reach it, and who should not
- A token reaches only its own tenant, within the scopes it was issued with, until it is revoked or its owner loses access.
- Test case and its limits
- Your developer runs a short script on staging that creates a read-only token for a-staff, calls a Tenant B record by id, calls a write route, revokes the token and calls again, without printing or saving the token's value. A second run, after a-owner invites a-leaver again at the staff role, creates the token for a-leaver and removes a-leaver instead of revoking. Limits: staging only, no more than ten requests per run. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- The Tenant B call and the write call are refused, and the token stops working once revoked. Whether it stops when its creator is removed is your rule: record it. The Multi-Tenant cheat sheet asks for API credentials bound to explicit tenant sets, environments and permission scopes.
- Evidence to keep
- The script's output of response statuses and the token's id, never its value, and the revocation time.
- What this row cannot show
- How tokens are stored. This part only reads an answer: ask your developer in writing whether each token is returned once and stored only as a hash, as the cheat sheet recommends for tenant API credentials.
- Buyer-audit fields
- Scope: the token routes and the API routes called, as a-staff, a-owner and a-leaver. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G11. Webhooks sent to customers
- Data involved
- Event payloads carrying a tenant's records, sent to the addresses that tenant configured.
- Who should reach it, and who should not
- Each tenant's events go only to that tenant's own endpoints. Only the roles your rules allow can add or change an endpoint.
- Test case and its limits
- Your developer runs a short script on staging that, as b-owner, registers a webhook to a receiving endpoint on a domain you control, without printing or saving any signing secret the product returns. Create one record in Tenant A and one in Tenant B. With the same script, as a-staff, try to add or edit Tenant A's endpoint if only owners may. Limits: staging only, your own receiving endpoint only, no more than six events. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- B's endpoint receives B's event and nothing from Tenant A, and the edit is refused where your rules say so.
- Evidence to keep
- The events your endpoint received, with their tenant markers, and the edit response.
- What this row cannot show
- Whether the address field refuses internal destinations, which the SSRF guide's allowlist verification table covers, and whether customers check your signatures.
- Buyer-audit fields
- Scope: the webhook settings routes and the delivery path, as b-owner and a-staff. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G12. Billing events received from a payment provider
- Data involved
- Subscription events naming a customer or subscription that your product maps to a tenant, which change its plan, limits and access.
- Who should reach it, and who should not
- A valid, signed event changes only the tenant mapped to that subscription. Nothing unsigned changes a plan.
- Test case and its limits
- Read the provider's terms first and use only the test mode it documents. In that test mode, change Tenant A's subscription and check Tenant B's plan. Send one unsigned event to the staging endpoint. Limits: staging and the provider's test mode only, no real card, no more than four events. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Only Tenant A's plan changes, and the unsigned event is rejected with no change to either tenant.
- Evidence to keep
- The provider's test event ids, your endpoint's responses, and both tenants' plans before and after.
- What this row cannot show
- The provider's own platform, which is not yours to test, and event types your handler has never received.
- Buyer-audit fields
- Scope: the incoming webhook route, in the provider's test mode. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G13. Subscription cancellation and offboarding
- Data involved
- Everything tied to the cancelled tenant: sessions, tokens, webhooks, file links, exports and stored records.
- Who should reach it, and who should not
- After cancellation, only what your policy allows, such as an owner downloading an export for a set period. Every other tenant is untouched.
- Test case and its limits
- First give Tenant C an invented file with a link, a webhook to your own receiving endpoint registered the way G11 does, and a token created for its owner the way the G10 script does, and give Tenant C-West one invented file. Cancel Tenant C on staging, never a real tenant. Then try Tenant C's open session, its token, the file link, a webhook event and a fresh sign-in. If storage keys use a tenant prefix, check that Tenant C-West's file survives. Limits: staging only, test tenants only, no more than fifteen requests, set-up included. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Your policy's result for each item, and nothing changed in Tenants A, B or C-West. The Multi-Tenant cheat sheet's offboarding steps include rejecting new operations, then revoking all active sessions, API keys and other tenant-scoped credentials.
- Evidence to keep
- Each response with its time, and Tenant C-West's file list before and after.
- What this row cannot show
- Whether data is later deleted from backups and replicas. G14 reads the policy for that.
- Buyer-audit fields
- Scope: the cancellation flow and each item tried afterwards, for Tenant C. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G14. Retention and deletion policy (read only)
- Data involved
- The written retention and deletion policy, and each place a tenant's data lives.
- Who should reach it, and who should not
- Your team and any adviser you choose. This row reads documents and asks questions, and sends no request.
- Test case and its limits
- Read only. List each place tenant data lives, using the Multi-Tenant cheat sheet's list of active stores, caches, object versions, replicas, exports, backups and encryption keys, and ask your developer in writing what happens to each after cancellation, and when.
- Expected result
- Every place has a written answer that matches the policy and what your customer contracts say. Where the Privacy Act applies to you, APP 11.2 may bear on this row; confirm that with your adviser.
- Evidence to keep
- The policy version, the developer's written answers and the date.
- What this row cannot show
- That deletion happened. The row records what is meant to happen.
- Buyer-audit fields
- Scope: the policy and each data store named. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
G15. Audit trail across the grid
- Data involved
- Audit records for invitations, role changes, impersonation, exports, token changes and cancellation.
- Who should reach it, and who should not
- Tenant owners read their own tenant's audit records, if your product offers them. Audit reads across tenants belong to your own staff.
- Test case and its limits
- Read the audit entries that rows G2 to G13 created. Then, as a-owner, open the tenant audit view, if there is one, and filter for B-CUSTOMER and for Tenant B users. Limits: staging only, no more than four requests. Test accounts and invented records only, no load, and any email or text goes to an address or number you control. Treat any link that works for whoever holds it as a secret: open it only for the invented record it belongs to, and never copy it into the evidence. Stop and record if a response shows a real person's data.
- Expected result
- Each action has an entry with the tenant, the actor and the time, and a-owner sees no Tenant B entries. The Multi-Tenant cheat sheet says tenant administrator rights must not grant platform-wide audit access.
- Evidence to keep
- The entries themselves, with any personal details masked.
- What this row cannot show
- Whether anyone reviews the records or whether alerts fire. The security logging and monitoring guide covers testing that an alert fires.
- Buyer-audit fields
- Scope: the audit records for G2 to G13 and the tenant audit view, as a-owner. Build and date: ____. Exceptions: ____. Remediation: ____. Retest status: ____.
What a SaaS security checklist has to prove
OWASP's Multi-Tenant Application Security Cheat Sheet describes multi-tenant applications as serving multiple customers, or tenants, from a shared infrastructure, codebase and often shared databases.
Each customer business you sell to is a tenant. Among the critical security challenges the cheat sheet says multi-tenancy introduces, a single vulnerability can expose all tenants' data. The security questionnaires guide lists the question that follows as a typical one: how do you ensure one customer cannot access another customer's data?
A SaaS security checklist worth running answers that question at each moment a tenant's data changes hands: when the tenant is created, when people are invited and given roles, when someone pays the bills but should not see the work, when your own support staff step in, when data leaves as an export, a file link, an API call or a webhook, and when the subscription ends.
Each moment is a separate way into the data, so this page treats SaaS application security moment by moment rather than screen by screen.
At Level 2, ASVS 5.0 requirement 8.4.1 asks for verification that multi-tenant applications use cross-tenant controls to ensure consumer operations will never affect tenants with which they do not have permissions to interact.
SaaS cyber security also covers your company's own accounts, laptops and suppliers. This page stays with the product your customers sign in to, where the tenants meet. It treats tenant isolation as a buyer-evidence problem as well as an engineering one, because a customer's questionnaire or an auditor can ask what was tested, what failed and whether it was fixed.
The grid on this page records those answers row by row. It is general information about testing, not legal or audit advice.
Where customer data moves, from sign-up to cancellation
These are the workflows the grid follows, in the order a tenant meets them.
For each one, the paragraph names the data that moves and the access question it raises. The grid turns each question into a test case.
A tenant can come into existence through self-serve sign-up, a setup your staff run, or an import. WSTG v4.2's registration test, WSTG-IDNT-02, asks among its questions whether users can register for different roles or permissions and whether the exchange of identity information can be manipulated during registration.
In a SaaS product, the sign-up request is also where a tenant id, a plan or an owner flag can be slipped in alongside the fields the form sends, and the mass assignment guide covers fields a client should not set.
The grid checks that each new tenant gets an identifier your server assigns and starts with no one else's records. The Multi-Tenant cheat sheet adds a rule for the provisioning code itself: treat onboarding as an all-or-nothing operation, so that if a step fails, partially created resources are removed and the failure is recorded.
An owner invites colleagues, and each accepted invitation attaches a person to a tenant at a role. That makes an invitation a provisioning step in its own right. WSTG-IDNT-03, the account provisioning test, asks whether an administrator or other user can provision accounts with privileges greater than their own, and whether there is any verification, vetting and authorisation of de-provisioning requests.
An invitation link may also let whoever holds it join, so the grid checks who can accept it, how many times, and whether withdrawing it stops it working.
Roles then decide what each person can do inside the tenant. WSTG-IDNT-01 asks the tester to review the granularity of the roles and the needs behind the permissions given, and a SaaS role list deserves that review before any test: can staff invite people, change roles, delete records or export everything?
Write the answers down as a permission matrix, a row for each function and a column for each role, before the grid runs.
Removal is part of the same workflow. ASVS 5.0 requirement 7.4.2, at Level 1, is that the application terminates all active sessions when a user account is disabled or deleted, such as an employee leaving the company. Removing someone from one tenant may not disable their account if they belong to others, so the grid checks the removed person's open session against the tenant they left.
When a session must end, on sign-out, on a timeout or when an account is disabled, is the session management guide's subject.
A finance contact who pays the invoices may need the billing pages and nothing else. If your product has that role, give it a column of its own in your permission matrix, and the grid tests it in row G5: it should reach the plan, the invoices and the billing contact, and it should not reach customer records, exports or member management.
Where card details stay with your payment provider, the role sees only what your product stores, and the provider's hosted pages sit outside the grid.
Your own support staff hold a role that can reach across tenants. OWASP's Session Management Cheat Sheet says web applications must thoroughly protect administrative interfaces that allow the management of all current active sessions, and that support personnel frequently use them to solve session-related or general issues by impersonating the user and looking at the application as the user does.
Whether your product allows impersonation, who may start it, whether the tenant is told, what it may change and how it ends are product decisions. The grid checks that the product does what you decided and that the record shows who really acted.
The OWASP Logging Cheat Sheet's list of events to log where possible includes use of higher-risk functionality, and names among it the use of systems administrative privileges or access by application administrators, including all actions by those users.
Data also leaves the product as an export file, a stored file served from a link, or a generated document. An export can build its own query rather than reuse the one behind the screen, which is why the broken access control and IDOR guide treats exports, and the background jobs that build them, as separate places for the check to go missing. Which columns an export should carry at all is the sensitive data exposure guide's subject.
Where files sit in object storage and are served through signed links, the Multi-Tenant cheat sheet asks for the exact tenant, object and operation to be authorised before a signed URL is generated, and for signed URLs to be restricted to the required object and operation, with an appropriate expiration and revocation model. The file upload security guide covers private storage and signed download links in more depth.
If customers connect your product to their other systems with API tokens, and your product tells their systems about events through webhooks, both carry tenant data outside a browser session, with no person on the screen to notice a mistake. The cheat sheet asks for API credentials to be bound to explicit tenant sets, environments and permission scopes, and for each tenant API credential to be returned once and stored only as a hash, never the plaintext value. The API security guide's endpoint register is the place to list every route a token can call.
Outgoing webhooks run the boundary the other way: each tenant's events should go only to the addresses that tenant configured. A webhook address is also a URL your server sends requests to on a customer's say-so, and OWASP's API7:2023 entry names webhooks among the features that encourage developers to access an external resource based on user input.
The SSRF guide covers that side. Incoming events, such as subscription changes from a payment provider, can change a tenant's plan and access, so the grid checks that a signed event changes only the tenant it maps to and that an unsigned one is refused.
When a subscription ends, the question becomes what still works. The cheat sheet's offboarding advice includes putting the tenant into a state that rejects new operations, then revoking all active sessions, API keys and other tenant-scoped credentials. It also asks for a documented retention and deletion policy applied to every place tenant data lives: active stores, caches, object versions, replicas, exports, backups and encryption keys.
Deletion has a boundary of its own. The cheat sheet asks for tenant deletion to target only that tenant's objects, using an exact prefix boundary. Its example is tenants/acme/ rather than tenants/acme, because the latter could also match tenants/acme-west/. Row G1 creates a second test tenant with a similar name, Tenant C-West, and the cancellation row checks that its file survives when Tenant C is cancelled.
Where tenant isolation breaks, and the guide for each failure
The grid does not re-explain each weakness.
When a row fails, the guide named beside it explains the cause and the fix. The common thread comes from the Multi-Tenant cheat sheet: client-supplied tenant identifiers should be treated as selectors only, with a check that the authenticated principal is authorised to act in the selected tenant.
- A tenant or record id taken from the request and trusted: the broken access control and IDOR guide, which has a two-account test matrix seeded across two tenants, and a lab that runs it.
- Tenant, role, plan or owner fields accepted at sign-up or in a profile update: the mass assignment guide.
- A session that outlives a role change, a removal or a cancellation: the session management guide.
- Exports, lists and background jobs that build their own query: the broken access control and IDOR guide. Exports that carry more columns than the task needs: the sensitive data exposure guide.
- File links that open for people who should no longer reach the file: the file upload security guide.
- API routes and tokens that no screen uses: the API security guide, starting from its endpoint register.
- Webhook addresses that send your server somewhere it should not go: the SSRF guide.
- Plan limits and downgrades enforced only in the screens: the business logic vulnerabilities guide.
- Support access and role changes with no record of who acted: the security logging and monitoring guide.
- Cached results shared between tenants: the Multi-Tenant cheat sheet asks for the tenant identifier in every cache key whose value or authorisation varies by tenant, and the grid has a row for search, counts and dashboards.
What buyers and auditors may ask a SaaS business to show
When a larger customer reviews a SaaS supplier, the questions may come in the customer's own spreadsheet or from a published set.
The Cloud Security Alliance says its CAIQ v4.1, released on 27 January 2026, offers an industry-accepted way to document what security controls exist in IaaS, PaaS and SaaS services. The security questionnaires guide covers how to answer a questionnaire with evidence. This section is about the tenant-isolation part, and the published lists that name it.
MVSP, the Minimum Viable Secure Product, describes itself as a list of essential application security controls and lists procurement among the places to use it. Its controls page, as read on 15 November 2026, includes these among the controls that bear on this grid:
- Control 1.2, customer testing: on request, enable customers or their delegates to test the security of the application, on a production or a non-production environment that closely resembles production in functionality. Ensure non-production environments do not contain production data. Reasonable restrictions on testing can be defined.
- Control 1.4, external testing: contract a security vendor to perform comprehensive penetration tests of the products, services and dependent systems at least annually.
- Control 2.7, logging: keep logs of authentication events (success and failure), create, read, update and delete operations on application and system users and objects, security relevant configuration changes (including disabling logging) and application owner access to customer data. Among its other requirements, logs must be stored for at least 30 days at no additional charge.
- Control 4.2, logical access: limit sensitive data access exclusively to users with a legitimate need, and the data owner must authorise such access. Deactivate redundant accounts and expired access grants in a timely manner, perform regular reviews of access to validate need to know, and ensure remote access to customer data or production systems requires multi-factor authentication.
Other buyer guides that reach the same rows
CISA's Secure by Demand Guide, as read on 15 November 2026, lays out questions and resources that organisations buying software can use to better understand a software manufacturer's approach to cybersecurity. It says cloud service providers and SaaS providers should retain and make available to customers security logs for at least six months at no additional charge, and it asks whether the manufacturer supports integrating standards-based single sign-on for customers at no additional cost.
The UK National Cyber Security Centre's supplier assurance questions, version 1.0, published on 17 December 2020, ask what provisions are in the contract for secure deletion or return of the customer's data and assets at contract exit, including transfer of services, data or assets to another supplier.
They also ask whether the supplier holds, or the customer would require it to hold, any cyber security certifications and, if so, whether their scope covers the parts of the service the customer will be using and how it will be using it.
The grid's cancellation and retention rows give you evidence for the contract-exit question. Each of these is a published list or question set, and each customer decides which, if any, it draws on.
SOC 2 reports and ISO/IEC 27001 certification
The AICPA's SOC suite page, as read on 15 November 2026, says System and Organization Controls (SOC) is a suite of service offerings CPAs may provide in connection with system-level controls of a service organisation or entity-level controls of other organisations.
The same page lists an AICPA guide on SOC 2 reporting on an examination of controls at a service organisation relevant to security, availability, processing integrity, confidentiality or privacy. On the Cloud Security Alliance's STAR Program page, as read on the same day, STAR Level 2 is a third-party audit: certification (ISO/IEC 27001), attestation (SOC 2) or C-STAR for the Greater China market, each combined with a CAIQ submission.
For this grid, what matters is who decides. A penetration test report or a completed grid is evidence that your SOC 2 auditor, or for ISO/IEC 27001 your certification body, may weigh. Whether it counts towards any control is their decision. Nothing on this page makes a product compliant or ready for an audit. Clever Ops is not certified to ISO/IEC 27001, holds no SOC 2 report and issues neither.
Australian privacy duties that may bear on a SaaS business
A tenant's records can include personal information about that customer's own clients and staff, which is where privacy law can meet product design.
In the Privacy Act 1988, as compiled in Compilation No. 104 with the text in force on 4 June 2026, section 6(1) says APP entity means an agency or organisation, and section 6D sets out what a small business and a small business operator are. Establish with your privacy adviser whether the Act applies to your business and to the personal information in your tenants' records.
Where it does apply, the grid's rows sit closest to Australian Privacy Principle 11, on security of personal information. APP 11.1 says that if an APP entity holds personal information, it must take such steps as are reasonable in the circumstances to protect the information from misuse, interference and loss, and from unauthorised access, modification or disclosure.
Under APP 11.3, such steps include technical and organisational measures, for the purposes of 11.1 and 11.2 and without limiting those subclauses or any other provision of the Act.
APP 11.2 applies where an APP entity holds personal information about an individual, no longer needs it for any purpose for which it may be used or disclosed by the entity under Schedule 1, the information is not contained in a Commonwealth record, and the entity is not required by or under an Australian law, or a court or tribunal order, to retain it.
In that case the entity must take such steps as are reasonable in the circumstances to destroy the information or to ensure that it is de-identified.
The grid's cancellation and retention rows record what your product does; what is reasonable for you is a question for your adviser.
The OAIC explains that under the Notifiable Data Breaches scheme, an organisation or agency covered by the Privacy Act 1988 must notify affected individuals and the OAIC when a data breach is likely to result in serious harm to a person whose personal information is involved.
Whether a cross-tenant finding is such a breach is a question to take to your privacy adviser. The broken access control and IDOR guide sets out what to do first if other customers' records were reachable.
Other provisions of the Act concern contracts with an agency, a term section 6(1) defines by a list that begins with a Minister and a Department. Subsection 95B(1) says section 95B requires an agency entering into a Commonwealth contract to take contractual measures to ensure that a contracted service provider for the contract does not do an act, or engage in a practice, that would breach an Australian Privacy Principle if done or engaged in by the agency.
Section 6D(4)(e) says an individual, body corporate, partnership, unincorporated association or trust is not a small business operator if it is a contracted service provider for a Commonwealth contract, whether or not a party to the contract.
If an agency is, or may become, a customer, read the privacy clauses of its contract with your adviser.
Running the two-tenant grid on your own product
Multi-tenant security testing needs two tenants that both belong to you.
Build them on a staging copy of your product, never on a real customer's tenant and never on a system you do not own or have written authority to test. The grid's description lists the invented tenants, users and records it expects.
OWASP's Authorization Regression Testing Cheat Sheet includes a cross-tenant boundary test: provision two tenants, seed data into one, run broad read queries as a user from the other, and assert that the response contains no records belonging to the first tenant. It says even a single leaked record identifier constitutes a critical failure. The grid applies that test to each lifecycle step, not only to the main screens.
- Name records so a leak is obvious. Every invented customer, file and export in Tenant A carries an A marker, and every one in Tenant B a B marker, so a B user who sees an A name has found a failure without reading further.
- Write the expected result for every row from your own rules before you send anything. The grid's expected results are a starting point. Change any that your product decides differently, such as records two tenants are meant to share.
- Keep the volume small: a handful of requests per row, no load, and nothing automated beyond the short scripts a row names. Every email or text a row triggers goes to a mailbox or number you control.
- Read the terms of each third-party service a row touches, such as a hosted sign-in service or a payment provider, before you start, and ask the vendor if they are unclear. Use the test mode a vendor documents where it offers one. The grid tests your product's side of each service, not the vendor's platform.
- Stop and record if any response shows a real person's data, rather than looking further. Record the request and what came back with the personal details masked, then follow the steps in the broken access control and IDOR guide.
- Run the grid again when a lifecycle step changes, such as a new role, a new kind of export or a new kind of token, and keep each completed copy with the build number it covered.
Filling in the buyer-audit fields
The last column of each row holds the fields a buyer or auditor may ask about. Scope names the routes, accounts, build and date the row covered. Exceptions records any result that differed from the expected one, and any part of the row that was not run, with the reason.
Remediation names the fix, who made it and the build it shipped in. Retest status takes one of four values: fixed, partially fixed, still reproducible or not retested.
A row that was not retested stays not retested rather than passed, so a later answer to a customer does not overstate what changed.
When a buyer asks how you keep one customer's data from another, send the scope and status fields rather than the reproduction steps. The security questionnaires guide covers what to share and how to word an answer that matches the evidence.
What a two-tenant review cannot show
A completed grid is narrow evidence.
Say what it covered and stop there, in the questionnaire as much as in the grid.
- It covers the steps, routes and accounts you listed, on the build you ran it against. A route nobody listed is untested, which is why the API security guide starts from a register of every endpoint.
- Staging is not production. Settings that differ, such as storage permissions or the sign-in service's configuration, can pass on one and fail on the other. Record the differences you know of in the scope field.
- Two tenants with a handful of records cannot show problems that need volume. The Multi-Tenant cheat sheet lists noisy neighbour attacks, one tenant exhausting shared resources and affecting others, among its key risks. Load testing is a job for a specialist, with your host's written approval.
- Records that tenants are meant to share, single sign-on through customers' own identity providers and complex multi-tenant billing need rows of their own before the grid says anything about them.
- Backups sit outside it. The cheat sheet notes that a shared backup can contain data from several tenants and describes a two-tenant restore test. The grid's retention row only reads your policy.
- A clean grid is not a certificate. OWASP's Authorization Cheat Sheet calls automated unit and integration testing of access control logic "not a substitution for a dedicated security test or penetration test", while saying such testing can help reduce the number of security flaws that make it into production.
Where a Clever Ops test fits
The grid is yours to run.
When a buyer or an auditor wants an outside view, SaaS penetration testing at Clever Ops is the web application and API penetration test: a signed-in test of a custom application that customers or staff sign in to, including the API behind it.
The scope asks for two test accounts in each role, each with sample records, so access between people in the same role can be tested in both directions, and for a product that serves separate customer organisations it asks for accounts in two of them as well.
Larger and multi-tenant applications are scoped individually after a longer call, and very logic-heavy products, such as marketplaces, lending platforms and multi-tenant billing, are scoped individually or referred to a larger firm.
Some parts of a SaaS product add effort and are scoped separately: teams with their own roles inside one customer account, records that customer organisations share with each other, and single sign-on through customers' own identity providers. A payment provider's hosted checkout and platform, and your hosting provider's platform, belong to those businesses and sit out of scope, as do third-party services on your subdomains unless their owner agrees.
On payments, the test stays with your application's side rather than the provider's platform.
An AI assistant built into the product is tested under the AI chatbot and agent security test, and a product made with an AI app builder is better served by the vibe-coded app security check.
Destructive actions are off by default for real people's accounts and records that existed before the test, and reset, lockout and deletion checks run only on test accounts and test records where the rules of engagement allow them. The report records the scope as tested, findings with redacted evidence, and negative results for the routes and accounts tried.
After the one consolidated retest, each finding is marked fixed, partially fixed, still reproducible or not retested, which are the same values as the grid's retest field.
A report is evidence for someone else's decision, not a certificate. Clever Ops is not certified to ISO/IEC 27001, holds no SOC 2 report and issues neither. Clever Ops is not CREST accredited, so where a buyer, panel or contract requires a CREST-accredited provider, we refer that work to a provider that is.
If a customer or auditor asks for an independent test, choose a provider that had no part in building or running the product: when Clever Ops built or hosts it, the work is a security review, and the report says who built it and who hosts it.
Frequently asked questions
For a multi-tenant product, start with tenant isolation across the account lifecycle: tenant creation, invitations, owner, staff and billing-only roles, support access, exports, file links, API tokens, webhooks and cancellation. Each is a separate way into a tenant's data, so each needs its own two-tenant test and its own evidence. Controls such as multi-factor sign-in, logging and backups sit beside those tests, and published lists such as MVSP set out controls a buyer may ask about.
Use a staging copy with two invented tenants, invented users in each role, and records whose names mark the tenant they belong to, so any leak is obvious. Write the expected result first, send no more requests than each row's limit allows, route every email to a mailbox you control, and stop and record if a real person's data appears. A third test tenant, created during the run, is the one you cancel.
No. In the AICPA's description, SOC services are offerings CPAs may provide, and the Cloud Security Alliance's STAR Program lists ISO/IEC 27001 certification among the forms of its Level 2 third-party audit. A test report is evidence your auditor or certification body may weigh, and whether it counts towards any control is their decision. Clever Ops is not certified to ISO/IEC 27001, holds no SOC 2 report and issues neither.
That is a product decision, and the grid tests whatever you decide. OWASP's Session Management Cheat Sheet says administrative interfaces that manage all current active sessions must be thoroughly protected, and that support personnel frequently use them to solve session-related or general issues by impersonating the user. If you allow it, decide who may start it, whether the tenant is told, what it may change and how it ends, then check that the audit record names the staff member as well as the user. MVSP control 2.7 includes application owner access to customer data among the events to log.
Your policy sets the timing. Among its offboarding steps, OWASP's Multi-Tenant Application Security Cheat Sheet says to put the tenant into a state that rejects new operations, then revoke all active sessions, API keys and other tenant-scoped credentials, and it asks for a documented retention and deletion policy applied to every place the tenant's data lives. Test it on a staging tenant made for the purpose, check that no other tenant is touched, and confirm any retention or deletion duty under privacy law with your adviser.
That is your decision, and the terms belong in your agreement with them. MVSP control 1.2 asks suppliers to enable customers or their delegates, on request, to test the application on production or on a non-production environment that closely resembles production in functionality, asks them to ensure non-production environments do not contain production data, and allows reasonable restrictions on testing to be defined. If you agree, give the customer a test tenant of their own and written rules, so their testing stays inside it.
Scope the right test for your application
Tell us what you have built, who uses it and when it needs to launch. You will get a plain-English recommendation: a website audit, a web application test or a referral to a specialist.