Security evidence
SaaS Due Diligence: What Proves Customer Data Is Isolated?
What SaaS buyers should request to assess tenant isolation: evidence, cross-customer access checks, coverage limits, and consequences for an acquisition decision.
By Andrey Shedko · · AI-assisted drafting
The investment question
A login screen and a tenant ID are not proof that customers cannot access each other’s data. Ask for evidence across the workflows that matter to the acquisition, and distinguish a demonstrated boundary failure from areas that were never tested.
1. Start with the exposure that matters to the buyer
A SaaS product may look credible in a demonstration while leaving a more consequential question unanswered: can one customer read or change another customer’s records? For a buyer, the issue is not whether the application uses a fashionable architecture. It is whether the customer separation assumed in the operating plan is supported by evidence.
Start with a short inventory of sensitive workflows: customer records, uploaded files, financial exports, search, integrations, and support access. Identify which would materially affect customer trust, business continuity, or the buyer’s ability to sell the product if separation failed. Do not convert that inventory into an unsupported estimate of breach cost or lost revenue.
AWS explicitly distinguishes tenant isolation from ordinary authentication. A successful login identifies a user; it does not, by itself, establish which customer’s resources that user may access. The diligence task is to connect the claimed boundary to the actual data paths. This article describes an evidence approach, not a security certification or a substitute for a scoped specialist assessment.
2. Ask for a traceable evidence chain
Request a short account of how the product determines the current customer, checks the user’s rights, and restricts the data returned or changed. Ask the reviewer to trace an important workflow from request to stored record and back. A diagram should help locate the control; it should not be the only evidence that the control exists.
OWASP’s API guidance calls for checking permission for the particular object and action. That makes a direct question useful: where is the decision enforced for this record and operation? A field named tenant_id, an unguessable record identifier, or a hidden menu is insufficient evidence on its own. Ask what enforces the rule when the normal interface is bypassed.
For every material conclusion, retain an evidence reference, reviewed revision, relevant configuration, test identity, environment, and date. Do not include production credentials or unnecessary customer data in the investor pack. A concise evidence index can point to controlled records held by the assessment team.
| Buyer question | Evidence to request | What remains unresolved without it |
|---|---|---|
| Where is the customer boundary enforced? | Reviewed access-control path and relevant deployed configuration for an agreed critical workflow. | Whether separation is a design intention or an implemented control. |
| Can an ordinary customer cross that boundary? | Authorized negative checks with two synthetic customers and explicit expected results. | Whether a valid login can access a record belonging to another customer. |
| Does the export follow the same rule? | Export initiation, job execution, generated file, and download authorization evidence. | Whether the customer-facing page and its background workflow have equivalent protection. |
| Who can bypass ordinary restrictions? | Support and service roles, approval arrangements, and relevant audit evidence. | Whether operational privileges create an unexamined route to customer data. |
| Will a repair remain effective? | A check that reproduces the observed failure before repair and passes after it, plus review of related paths. | Whether a local fix addresses the demonstrated condition and guards against recurrence. |
3. Define a small, meaningful demonstration
Agree written authorization, environment, permitted actions, data handling, and stop conditions before testing. Use synthetic customers and records wherever possible. The purchase process does not authorize probing a seller’s production service, accessing real customer records, or running destructive checks.
A useful demonstration has an allowed case and a forbidden case. First establish that the test user can perform the intended action on its own synthetic record. Then check that the same user cannot perform that action on the other synthetic customer’s record. Otherwise, an apparent denial may simply reflect a broken endpoint or invalid test setup.
Record the actual result, including relevant response content and side effects. A rejection status alone does not establish that no unauthorized change or background action occurred. Conversely, a successful HTTP status does not necessarily mean data was exposed: the response may contain a valid empty result. The reviewer needs to interpret the behavior against the agreed access policy.
Keep the result narrow. A successful check demonstrates the observed outcome for the chosen identities, operations, revision, and environment. It does not establish that every customer, role, endpoint, or deployment behaves the same way.
4. Look beyond the main customer screen
Choose coverage from the product’s important workflows rather than counting tested pages. A customer record may be correctly protected in the main screen while appearing in a generated export, search result, attachment download, or integration response. Ask which of those paths exist and how they were assessed.
OWASP’s multi-tenant guidance covers boundaries beyond database queries, including caches, asynchronous work, and file storage. For the buyer, the practical implication is to request a coverage map showing the relevant paths, the check performed, and the untested areas. Avoid treating a large endpoint count as a reliable measure of assessment quality.
- Exports and reports: follow the request through completion and download; checking only the button that starts the job leaves later stages unresolved.
- Files: distinguish permission to view an application record from permission to retrieve its attachment or an already issued download link.
- Background work: identify the customer context used when a queued task runs, particularly when the worker has broader privileges than an ordinary user.
- Support access: record intentional cross-customer capabilities separately from a customer-boundary defect. Assess who may use them and what evidence of use is retained.
- Membership changes: where material, check what happens after a user changes organizations or loses access; document any token or session assumptions.
5. Interpret a finding without inventing its history
Fictional illustration — not a client case or an observed production incident: a reviewer uses two synthetic customers in an authorized test environment. Ordinary record access is denied across the boundary. A completed export belonging to the second customer can nevertheless be retrieved using the first customer’s session. The selected export check returns the other synthetic customer’s data.
The supported finding is an authorization failure on the tested export path. It does not prove that every data path is vulnerable, that real customer data was accessed historically, or that a reportable breach occurred. It also cannot be dismissed by pointing to the successful record-access check: the export is a separate demonstrated route.
The buyer should ask whether the affected revision and configuration represent the proposed acquisition environment, what other exports share the relevant control, and what containment is feasible. Historical exposure is a separate investigation. Logs may help, but incomplete retention or missing detail can prevent a confident conclusion. 'No evidence found' must not silently become 'no exposure occurred'.
6. Translate the evidence into an acquisition decision
The report should distinguish a confirmed boundary failure, a plausible concern awaiting evidence, and an area excluded from review. These states support different decisions. A confirmed failure on a sensitive workflow may require containment and retesting before the buyer relies on that workflow. An untested path calls for additional work or explicit acceptance of uncertainty; it is not a clean result.
Ask for a bounded remediation plan: affected paths, proposed control, investigation of similar paths, regression checks, release validation, owner, and dependencies. Do not assume that adding one missing filter resolves every exposure. Equally, do not recommend a platform rewrite without evidence that a narrower repair cannot meet the requirement.
The investment committee needs the consequence, the basis, and the remaining uncertainty. State which planned use of the product is constrained, what evidence would remove that constraint, and which decision still requires specialist input. Potential contractual, privacy, or notification implications should be assessed by the relevant advisers using the actual facts, rather than asserted from a technical test alone.
- Before reliance: identify any demonstrated sensitive-data path that needs containment or validated repair, and the operating assumption affected.
- Further diligence: name missing access, environment differences, untested roles, and workflow coverage that could change the conclusion.
- After repair: retain reproduction evidence, regression results, deployment confirmation, and any unresolved investigation limits.
- For AI-built products: apply the same evidence standard. AI-built describes development with coding tools; AI-powered describes runtime use of models. Neither label proves or disproves tenant isolation.
Sources & further reading
- AWS SaaS Architecture Fundamentals: Tenant isolation ↗
Primary guidance distinguishing tenant isolation from basic authentication and explaining the role of tenant context. It does not establish the security of a particular application.
- OWASP API Security Top 10: Broken Object Level Authorization ↗
Primary guidance on authorizing access to individual objects and actions. The buyer-facing evidence and decision framework in this article is editorial guidance.
- OWASP Multi-Tenant Application Security Cheat Sheet ↗
Primary implementation guidance covering tenant context and boundaries across application components. Linked as a technical reference, not a completed assessment or certification.
The checklists and decision framework are Shedko Advisory editorial guidance, informed by the linked technical references. References do not imply endorsement, certification, or a completed client assessment. Examples labelled fictional are illustrative; conclusions in a real engagement depend on its scope and evidence.
Apply this to your acquisition
Define the decision. Then agree the assessment.
Discuss technical due diligence for a SaaS acquisition →
Explore the clearly fictional sample report →
Request a Confidential Assessment ↗Start with a non-confidential overview. Practice in preparation · preliminary inquiries welcome.