Secure AIProof, not promises.
Try it yourself, and see how often it catches each kind of private detail.
84%of private details caught on our hardest test
100%of card numbers, phone numbers and emails
0messages ever used to train an AI
Try it
Use made-up details. See exactly what the AI receives, then the answer with your details put back.
Use made-up details. Nothing you type here is saved.
How often it catches each detail
84% of 50,187 labelled private details across 31,001 documents nothing was tuned on.
Card numbersEvery card in 2,500 labelled documents100%
IBANsThe official example for all 121 IBAN countries100%
Phone numbersThe official example for 245 countries and territories100%
Email addressesEvery email in the labelled test documents100%
National ID numbers1,022 numbers from 43 countries, when the text says what they are99%
Passwords and API keys5,963 example secrets from gitleaks, an open-source scanner97%
PostcodesGoogle's example postcodes for 169 countries96%
People's names31,001 documents nothing was tuned on93%
Street addressesMicrosoft's synthetic test data, in dozens of countries' formats89%
Social Security numbers2,500 labelled documents89%
Account and reference numbers with no label8,857 documents with such numbers81%
Bank account numbers12,500 labelled documents70%
How every test was run is on Benchmarks.
What we keep
What the AI provider getsThe question, with your private details removed. Never your name, your account, or who is asking.
TrainingNothing you write trains any AI, ours or a provider's. It's in our contract with every provider.
Your conversations and filesStored encrypted. Staff only open them when their work requires it.
Quality recordsPrivate details already removed. Deleted automatically after 90 days.
Passwords and cardsNone. You sign in with Apple, Google or Microsoft, and Stripe handles payments.
When you leaveYour data is deleted within 30 days, except what the law makes us keep, like billing.
Everything we can open is on Security. The legal terms are in the Data Processing Addendum, and the companies we use are on sub-processors.
Check it yourself
The sealed serverOn Ultra, your message is sealed on your device and opened only inside an AWS Nitro Enclave we cannot see into. Your browser checks this fingerprint, signed by Amazon's hardware, before it sends anything.
58ecfbe30ed3dadcbce4b1b4da9de629de3a7b12c60c7aea42dc95abd0e100d8f9beb746541c0eec6fabdb6e70b77effPrevious version, kept for rollback: 485083262e3f5379…Outside checksOpen source. Our developer kit is on GitHub: secure-ai-sdk.Found a problem? Email company@secureai.one. We reply within 3 working days and never take legal action over a good-faith report.