Security questionnaire answers
Last updated: 8 November 2026
The 38 questions most security questionnaires ask, with our answers: 23 yes, 10 in progress, 5 no. Where the answer is no, it says what is planned. Every answer matches our security fact sheet. Send your own questionnaire to hello@cleverfox-ai.com and we will fill it in from these. Back to the Trust centre.
Governance
Yes. A 40-point security fact sheet at /security, each check marked in place, in progress or not yet applicable and verified against the running code; a privacy policy at /privacy; terms of service at /terms.
No. We do not hold a certification yet.
Next: Cyber Essentials is next on our roadmap, then ISO 27001.
No, not yet. Our own automated checks run on every build: a census that fails the build if an endpoint has no access check, and a dependency scan.
Next: Not yet booked. Ask us if you need one before you sign and we will tell you where it stands.
Yes. Email hello@cleverfox-ai.com with SECURITY in the subject. A person replies within 2 business days, and a confirmed finding gets a fix or a plan within 14. There is a security.txt file at /.well-known/security.txt.
Access control
Yes, by two independent layers. The application checks every request for who you are and which company you are in, and the database enforces workspace-scoped row policies, so it refuses another company's records even if the application got it wrong.
Yes. Every endpoint is gated server-side, roles are checked on each request and never trusted from the browser, and an automated census fails our build if a new endpoint ships without an access check.
Partly. The database layer was tested live on 19 August 2026 with a simulated outside account, which saw zero rows. A full per-endpoint adversarial pass is the remaining step.
Yes. A small number of named Cleverfox administrators have access, for support. Every administrative action and data export is written to a tamper-resistant audit log that its own administrators cannot edit. We do not yet notify a customer each time an administrator accesses their data.
Not yet. Two-step sign-in for administrators is being rolled out now.
Next: Administrators first; we will update this answer when it is enforced.
No, not yet.
Next: Administrators come first. Extending it to every user is not yet scheduled.
Passwords are hashed by our authentication provider, never stored by us in readable form. One minimum-length rule applies on every path that sets a password, and sign-in and password reset are rate-limited per account.
Sessions are cookie-based, with no tokens in browser storage. Removing a person's account ends all of their sessions, and a written offboarding runbook was exercised for real in August 2026.
No.
Next: Not scheduled. Tell us if your organisation needs it.
Encryption and keys
Yes. In transit over HTTPS, with browsers told never to fall back to an unencrypted connection (HSTS). At rest, by the database provider. The off-site backup bucket is encrypted at rest too.
No credentials are in source control, and the full history has been swept. Privileged keys live only in the deployment platform and never reach the browser. There is a written access inventory and a rotation runbook for every credential; a suspected leak means rotating the same day.
Application security
No raw SQL: every query is parameterised. Request bodies are schema-validated with hard limits. All rendered rich text, including AI output, passes through a sanitiser, backed by a Content-Security-Policy. Cross-origin access is strict same-origin.
Yes. The production dependency tree is checked on every audit pass, with zero known vulnerabilities at the last check.
Yes. Content-Security-Policy, HSTS, clickjacking protection, strict referrer and permissions policies. No stack traces reach the browser, and a failed sign-in gives one constant answer so nobody can probe which accounts exist.
Uploads are authenticated, scoped to the workspace and size-capped, and active-content file types are refused. Some refinements are documented decisions still under review.
Logging
Yes, since 19 August 2026, in a tamper-resistant log that its own administrators cannot edit.
Yes. Every call our servers make to an AI provider writes a log row: which provider, which workspace, what happened. The log never keeps what was said. The one exception is a live voice conversation in The Room, where your browser talks to ElevenLabs directly.
Backups and continuity
Yes. The database provider takes a daily backup, kept for seven days, in London. Every night a second copy of the database goes to a separate provider in Amsterdam, encrypted at rest and locked against deletion for 30 days. Uploaded files are not yet in the nightly off-site copy; they stay with the database provider.
A quarterly restore-test procedure is written and the first dated test is scheduled. We do not publish recovery targets yet.
Next: Recovery targets follow the first dated restore test.
No. We have no status page of our own and no uptime commitment. Our hosting and database providers publish theirs, and /trust links to both.
Next: A status page of our own is not built yet.
Incident response
Partly. The audit log lets us establish what was accessed and when, and every credential has a written same-day rotation procedure. Formally naming who leads a breach response is in progress.
Without undue delay. Our draft DPA commits to telling you within 48 hours of our becoming aware of a breach of your data, and where the law requires it we report to the Information Commissioner's Office within 72 hours.
Data location and sub-processors
Stored in London, UK. The application servers run in London, UK. The nightly off-site backup is in Amsterdam, Netherlands. AI providers process requests in their own data centres, mostly in the US, and only when you use an AI feature.
The Netherlands is in the EU, which UK law recognises as adequate. For providers outside the UK and EU, mainly the AI providers, we are confirming each provider's transfer terms, such as the UK International Data Transfer Addendum.
Yes. The list is on /trust with what each receives, where it is and whether we have finished checking its terms. We will email workspace administrators at least 30 days before a new sub-processor starts receiving customer data.
Not all on file yet. For most providers the standard DPA is part of their terms; we are locating, accepting and filing each one.
Yes. Our DPA is published at /trust/dpa. The current version is a template under review by our solicitor; ask us for a countersigned copy.
Retention and deletion
Yes. Workspace content is kept for the life of the workspace or until you delete it; AI usage logs and the audit log for 24 months; onboarding enquiries that go nowhere for 12 months. The full schedule is available on request.
Yes. Deleting a person removes their sign-in, their uploaded files and their identity from team content, with a report of each step recorded in the audit log. If your company leaves, you have 30 days to export everything, then we delete it from our live systems. Copies in backups are not restored or used; the off-site copies are locked for 30 days, and an automatic expiry after that is not yet set up.
Yes. Export is built in, and you can ask us for a machine-readable export of your data at any time.
AI
We do not train AI models on your data. Anthropic, our main AI provider, may not train on customer content under its commercial terms (confirmed 2 November 2026). OpenAI's stated default is no training on API content. For OpenRouter, Deepgram, TypeSafe and ElevenLabs we are still verifying the position.
Yes. Every AI provider is named on /privacy and /trust with what it receives.
Outside content entering a prompt, such as a transcript or an uploaded document, is fenced as evidence, not instructions. Extending this to every fox is scheduled.
Yes. AI in the product is always presented as AI, and the privacy policy says plainly that AI output should be reviewed. A safety floor rides on every model call: a person in crisis gets real support routes, enforced on the server and tested on every build.