Security
Security, data hosting and evidence
Database and sign-in in Frankfurt. Separation of customers by database rules. No training of a language model on contract content.
No certification. Self-assessment against ISO/IEC 27001:2022 on 16 November 2026.
As of 18 November 2026
Overview
| Subject | Status | Basis |
|---|---|---|
| Database and sign-in | Supabase, Frankfurt (AWS eu-central-1), since 19 September 2026 | Project configuration |
| Contract data in the delivered files | none; the application loads data only after sign-in | Configuration |
| Encryption in transit | HTTPS on all addresses, HSTS valid for one year | Measured 17 November 2026 |
| Encryption at rest | by the database provider | Provider statement |
| Separation of customers | Tenant key and database rule on 39 tables | Test on 22 September 2026 on a copy |
| Training on contract content | excluded | Commercial terms of the model provider |
| Two-factor sign-in in CLEVER | not available; sign-in by link or with a Google or Microsoft account | Status of the application |
| Certification | none; no ISO/IEC 27001 certificate, no SOC 2 report | — |
| Reporting of vulnerabilities | info@clever-ai.org, /.well-known/security.txt | RFC 9116 |
Hosting and data storage
- Database and sign-in are hosted by Supabase in Frankfurt (AWS eu-central-1). Moved from Ireland on 19 September 2026.
- Website and application are delivered through Netlify. The delivered files contain no contract data; the application loads data from the database after sign-in.
- For the analysis, the contract text passes through a server function to the language model. The server function stores nothing.
- Stored are the contract text read from the document, the file name and the analysis. The original file is not stored.
- Internal paths for documentation, database changes, scripts and tests answer with status 404 on the website.
Encryption and response headers
- HTTPS on all addresses. HSTS with max-age=31536000, measured on 17 November 2026.
- Data at rest is encrypted by the database provider (provider statement).
- The key to the language model interface is held on the server only and does not reach the browser.
- Every response of the website carries X-Frame-Options SAMEORIGIN, Content-Security-Policy frame-ancestors 'self' and X-Content-Type-Options nosniff. Embedding in third-party pages is blocked.
- The TLS version is not measured and is therefore not stated here.
Access, roles and tenant separation
Sign-in
Link by e-mail or a Google or Microsoft account. CLEVER stores no password. A new account can be used only after approval. Two-factor sign-in in CLEVER: not available.
Roles
Rights per role, among them Admin, Deal Manager, Risk Manager and Approver. Enforced by database rules per role (Row Level Security), also on direct access to the interface.
Tenants
Tenant key on 39 tables, set by the database; a foreign value sent by the browser is rejected. The members are decided by the tenant’s Admin. The operator creates tenants, adds no members and assigns no roles.
Test of the separation
On 22 September 2026 on a copy of the database schema: 25 of 25 checks without a foreign row, 195 of 195 pairs of role and table unchanged. Internal test, no third-party audit.
Residual risk
One database rule separates the customers of a shared database; an error in it would affect all tenants. An own database per customer is available by agreement.
Log
Deletions, restorations and permanent deletions are recorded in a log that only appends entries. Read access for the Admin only.
Sub-processors
| Service | Purpose | Contract data | Location | Evidence per provider |
|---|---|---|---|---|
| Supabase | Database and sign-in | yes | Frankfurt (AWS eu-central-1) | ISO/IEC 27001:2022, SOC 2 Type 2 |
| Netlify | Delivery of the website and the application, server functions | in transit only, not stored | Edge nodes without assured location; server functions USA | ISO/IEC 27001, SOC 2 Type 2 |
| Anthropic (Anthropic Ireland, Limited) | Analysis by the language model | yes | outside the EEA | SOC 2 Type 2, ISO/IEC 42001 |
| Amazon Web Services (Bedrock) | Analysis in the EU, only with the add-on “EU processing” | yes, with the add-on | Frankfurt | ISO/IEC 27001, BSI C5 |
| Resend | Delivery of notifications and of the sign-in link | partly: subject, recipient, content of the notification | USA | SOC 2 Type 2 |
| Sign-in with a Google account; reading a document selected in Drive | only on selection by the user | EU and USA | ISO/IEC 27001 | |
| Microsoft (Microsoft Ireland Operations Ltd.) | Sign-in with a Microsoft account (Entra ID) | none | EU (EU Data Boundary); USA in individual cases | ISO/IEC 27001 |
Without contract data, with personal data: Microsoft 365 (e-mail and appointments); GitHub (source code and automated checks with test accounts and synthetic contracts, started hourly by a server function; in reports on failed analyses a pseudonym of the user instead of the e-mail address); Plausible (visitor statistics of the website, not in the application).
Without personal data: Cloudflare (DNS without routing through Cloudflare); European Central Bank (exchange rates, retrieved by the server).
Independent controller: Paddle (payment processing as reseller; the application loads the Paddle checkout only on purchase).
Evidence as stated by the providers, retrieved on 16 November 2026. Anthropic’s sub-processors are listed at anthropic.com/subprocessors. Legal basis of transfers: privacy notice, section 4.
Language model
- CLEVER trains no model on customer data. Under its commercial terms, the provider of the language model may not train on customer content.
- The provider deletes inputs and outputs within 30 days, except where required by law or for misuse review.
- With the add-on “EU processing”, the analysis runs through Amazon Bedrock in Frankfurt.
- Prepared, not active: Mistral AI (Paris, processing in the EU) and STACKIT (Schwarz Digits, data centres in Germany) as further language-model providers. No calls, no contract data. Listed as sub-processors before first use.
Reporting of vulnerabilities
Address
info@clever-ai.org, subject “Security”. Machine-readable at /.well-known/security.txt (RFC 9116).
Content
Affected address or function, steps to reproduce, time, own contact address.
Limits
Tests with an own account only. No access to data of other customers beyond the proof, no modification or deletion of data, no load or denial-of-service attempts, no deception of people.
Response time
First response within 5 business days.
Confidentiality
No publication before the fix.
Rewards
There is no bounty programme.
Security incident
- Reporting of an incident or suspicion by customers: info@clever-ai.org, subject “Incident”.
- Obligation as processor: notification of a personal data breach to the customer as controller without undue delay (Art. 33(2) GDPR).
- Written incident process with classification, roles and deadlines: draft of 17 November 2026, approval by the management pending.
ISO/IEC 27001
Certification
None. No ISO/IEC 27001 certificate, no SOC 2 report, no BSI C5 attestation.
Self-assessment
On 16 November 2026 against ISO/IEC 27001:2022: clauses 4 to 10 and the 93 controls of Annex A, each with evidence or gap. No third-party audit.
Management system
Drafts of 17 November 2026: scope, information security policy, roles, risk assessment method, inventory, access policy, incident policy, endpoint policy. Approval by the management pending.
Audit
No date agreed with a certification body.
Questionnaires and contracts
Supplier questionnaire and data processing agreement on request via info@clever-ai.org.
Documents
Security questions
Supplier questionnaires, data processing and architecture: info@clever-ai.org.