Start with one important decision
Use this kit before launch, a team handover or a repair-or-rebuild decision. It works for AI-assisted, contractor-built and in-house apps. Start with the brief, then choose the checks relevant to your product. Give the technical checks to your engineer if needed; the deeper checks can take more time.
Use only systems you own or are authorised to test. Hands-on checks belong in local or staging environments with dummy records and test accounts. Do not use customer data, live payments or destructive actions. If you cannot check safely, record Unknown. This kit is not a security certificate.
APP / DATE:
DECISION AND DEADLINE:
WHAT THE APP DOES / STACK, IF KNOWN:
USERS AND ROLES:
MOST IMPORTANT CUSTOMER JOURNEY:
WHAT MUST NEVER HAPPEN IN THAT JOURNEY:
SENSITIVE DATA, PAYMENTS, FILES OR AI ACTIONS:
KNOWN PROBLEMS:
WHO CAN EXPLAIN AND DEPLOY THE CODE:
Example rule: a member may download their workspace's documents, but never another workspace's documents. That is checkable. 'We have authentication' describes a feature, not its boundaries.
Record evidence, not ticks
- Evidence recorded: name the path, configuration or test result, revision, date and environment. Say exactly what it establishes.
- Concern observed: record what differs from the intended rule, without guessing the full impact.
- Unknown: name the missing evidence and who could help. An unknown is not proof of a vulnerability.
- Not applicable: explain why this feature or risk is absent.
Use these labels for each of the seven checks. One successful test covers that path at that revision. A hidden button, an AI assurance or a clean scanner result does not establish the whole app is safe.
1. The two-account check
Can one account reach another account's private records? Write the intended rule first. Create dummy accounts A and B in separate workspaces (or your actual ownership model). As A, create a private test item and record its known test URL or identifier.
In a separate browser profile signed in as B, request A's known dummy item directly. Ask an engineer to check the underlying API/download path too, including database and storage rules. Test signed-out access separately. Do not guess identifiers belonging to real users.
Expected evidence: A's permitted access works; B and signed-out requests return no private data. Record responses and confirm denied writes leave state unchanged. A redirect alone does not prove the API refused access. If private data is returned, involve the responsible engineer promptly to contain and investigate the exposure before expanding use.
2. Privileged credentials
Ask an engineer to distinguish public client keys from privileged credentials, then inspect browser-delivered code, repository history and relevant logs. Supabase publishable/legacy anon keys are designed for clients; secret/service-role keys are privileged. An .env filename alone does not establish that a build keeps a value private.
Optional engineer check, from a local Git repository after installing Gitleaks: the command below scans Git history and redacts detected secrets in output. It does not cover every exposure location. Keep tool version, coverage and redacted observations, never key values. If a privileged key was exposed, have its owner revoke or rotate it and assess access; deleting the latest copy is insufficient.
gitleaks git --redact
3. Paid access and repeated events
Skip with a reason if there are no paid entitlements. In the payment provider's test environment, ask your engineer to check that visiting a success URL or submitting a browser 'paid' value cannot grant access. The server must establish entitlement from trusted payment state.
Ask them to verify webhook signatures using the provider's supported method, replay the same test event and check late or out-of-order events. Expected evidence: the intended business effect occurs once, invalid events are refused and delayed events do not incorrectly change access. Keep the handler, signature-check location and observed results. A search for one SDK function is not proof of correct verification.
4. Restricted actions without the button
Choose one important action: changing roles, inviting a teammate, exporting data or deleting a dummy item. Write who may perform it. Have an engineer call its server/API path with a lower-privilege test account and while signed out, then with the intended role.
Expected evidence: the permitted action works and unauthorised requests neither disclose data nor change state. Keep the role, path, response and before/after state. Include direct database/storage access if the app exposes those paths. Hiding the menu is insufficient. If the rule fails, investigate its shared permission layer and related actions.
5. AI tools, data and spending
Skip if there is no AI feature. List the data it retrieves, tools it calls and actions they allow. Ask which permissions, usage limits and approvals the application enforces independently of model output.
In an isolated sandbox, ask the engineer to test a request for another dummy account's data or an action the test user may not authorise. Inspect configured per-user limits and failure behaviour using a small agreed test budget. Do not repeatedly hammer live login, email or paid AI endpoints. Expected evidence: the application refuses forbidden data/actions and enforces the intended usage limit. A prompt asking the model to behave is insufficient.
6. The next engineer check
Can someone other than the builder follow setup instructions and run the important flow with dummy data? Review repository scripts before running them. Ask what the relevant tests assert: permitted behaviour, denied access and an important failure or retry. Keep setup gaps, test names/results and the revision they cover.
For npm projects with a package lock, this optional command checks dependencies while omitting dev dependencies from the submitted audit. It sends dependency information to the configured registry; use it only where that is permitted. It does not change installed versions. Review advisories, installed versions and affected usage with an engineer. Development/build dependencies still need a separate review. Do not blindly apply audit fix commands or bulk upgrades.
npm audit --omit=dev
Expected evidence: another engineer can run the selected flow and explain what the tests and unresolved warnings establish. A dependency warning is a lead to investigate, not automatically an exploitable problem in your app.
7. The bad-day check
Name who receives failure alerts, controls deployment and can restore the needed data. Ask for the last relevant restore-test evidence, rollback instructions and account ownership/access arrangements. Database backups may exclude uploads and external-service state.
Expected evidence: named owners, what is backed up, a dated recovery exercise and known gaps. 'Backups enabled' is a setting, not a demonstrated restore. Plan any recovery exercise in isolation; never restore over production for this kit. Missing evidence becomes an owned investigation task.
Choose your next three actions
- Observed private-data access, exposed privileged credentials or a similarly serious failure: involve the responsible owner promptly, agree containment and investigate. This kit cannot establish whether real data was accessed.
- Missing evidence on an important permission, payment or data flow: make verification a task before relying on it. Do not turn every unknown into a critical finding.
- Other improvements: prioritise demonstrated impact on the next milestone. Keep lower-value cleanup separate. Avoid calendar deadlines based only on a scanner score.
ACTION / OWNER / REVIEW DATE:
OBSERVATION OR UNKNOWN:
EVIDENCE / REVISION / ENVIRONMENT:
IMPACT ON OUR DECISION:
NEXT CHECK OR CHANGE:
ACCEPTANCE CHECKS:
Worked finding — illustrative only
Example only, not a client result: an export handler checks sign-in, then reads an export by ID. The traced source path contains no workspace-membership check. This supports a concern; runtime exposure and upstream controls have not been verified. No severity is assigned without that context.
Next action: trace the full access path and run the authorised two-workspace dummy-data test. If confirmed, enforce membership before returning data and inspect related downloads. Acceptance: permitted exports work, cross-workspace and signed-out requests disclose no private data, and the intended revision/environment is recorded. State: open, not retested. Estimate after tracing the shared access layer.
Hand this brief to any engineer
WE NEED TO DECIDE / BY WHEN:
PRODUCT / STACK:
IMPORTANT FLOW AND PERMISSION RULE:
TOP THREE CONCERNS OR UNKNOWNS:
EVIDENCE AVAILABLE (NO SECRETS):
WHAT HAS NOT BEEN CHECKED:
TECHNICAL CONTACT:
HELP NEEDED: investigate / review / repair / handover
Your existing developer may be able to resolve the questions. If you want an independent investigation, Ekyon's focused audit starts at US$1,000: up to nine assessment hours including setup, investigation and reporting, plus a separate one-hour walkthrough. Scope is agreed first. Larger reviews, repairs and follow-up verification are separately scoped.
To discuss fit, reply to this email (hello@ekyon.io) with your product, stack and decision. Leave out passwords, key values, repository access and customer data. Confidentiality and access are agreed before code is shared.
See the audit scope and illustrative report
Ekyon, Code Audit Starter Kit · 12 November 2026. Keep the evidence and use the brief with any engineer.
Reference notes (commands checked against official docs, 12 November 2026)