Every AI vendor has a security page and every security page says the same things. These are the questions where the answers genuinely differ — and four where a smooth answer is the warning sign.
About the data itself
- •What exactly leaves our environment, and at what point? Not “your data is secure” — which fields, at which step, to which party.
- •Which sub-processors see it? Most AI vendors are a thin layer over somebody else's model. That somebody is in scope for your assessment whether or not they are named on the homepage.
- •Is it used for training, by you or by your sub-processors? Ask about both. A vendor can honestly say they do not train on it while the model provider behind them does.
- •How long is it retained, and what triggers deletion? “Until you delete it” is not a retention policy. Ask what happens to logs, backups and abuse-monitoring copies.
About what happens when it goes wrong
- •Who is the controller and who is the processor, in writing? If the vendor is vague here they have not done the work, and you will discover it during an incident.
- •What is the notification window for a breach, in hours? A contract that says “without undue delay” and nothing else has said nothing.
- •Can you produce a record of what was sent, for a given user, on a given day? This is the question that decides whether you can answer a subject access request.
- •What happens to our data if you are acquired, or you shut down?
The four where a confident answer should worry you
Some questions have no clean answer. A vendor who gives one has either not understood the question or is managing you.
- •“Can you guarantee the model will never output our data?” The honest answer involves the word probably. Anyone who says never is describing a system nobody has built.
- •“Is your redaction 100% accurate?” No detection is. Ask instead what happens when it is not — whether the request proceeds or stops, and who chooses.
- •“Are you GDPR compliant?” Compliance is a property of a use, not of a product. The right answer is a description of what they provide towards it.
- •“Can you delete everything about us on request?” Ask how, and whether that includes backups and any derived artefacts. Confident yes without detail usually means nobody has tried.
A vendor with no caveats has not thought about it, or is not telling you.
Our own answers
Ours, in the same terms. Secure AI reads files on the device rather than uploading them, and replaces every identifying detail before anything reaches a model. The mapping between a stand-in and a real person is never stored — so nobody can turn a redacted document back into real people, including us, and including anybody who asks us to.
Where detection cannot reach a confident answer, the request stops rather than proceeding — a strict mode the customer controls, because whether to send or to refuse is a decision that belongs to them.
Try Secure AI free
Frontier AI chat — encrypted by default, with your identity stripped before any request reaches the provider.
Common questions
What should I ask an AI vendor about data retention?
Ask what triggers deletion and what it covers. “Until you delete it” only describes the record you can see; logs, backups and abuse-monitoring copies usually have their own schedules, and those are the ones that matter during an incident or a subject access request.
How do I find out who really processes our data?
Ask for the sub-processor list in writing. Most AI products are a layer over somebody else's model, and that provider is in scope for your assessment whether or not they appear on the vendor's homepage.
Is 'we are GDPR compliant' a good answer?
No. Compliance is a property of a specific use, not of a product. A useful answer describes what the vendor provides towards it — a processing agreement, a transfer mechanism, deletion capability, a sub-processor list — and leaves your obligations to you.