Security
How Secure AI protects your identity, your messages and your data.
Secure AI is designed so that AI providers never learn who is asking. Private details are removed before a message leaves, what you save is stored encrypted, and nothing you write is used to train AI. This page describes how each of those works, and exactly what Secure AI itself is able to access.
For the reasoning behind the design, see Privacy. For the legal terms, see the Privacy Policy. The companies that process data for Secure AI are listed under sub-processors.
Account security
Accounts are created and accessed through Sign in with Apple, Google or Microsoft, or with a sign-in link sent by email. Secure AI never receives or stores a password, so there are no passwords to steal. Each account can access only its own conversations and documents.
Message protection
Every message is sent to Secure AI first, never directly to an AI provider. Before it is forwarded, names, account numbers, addresses and other private details are removed, so the provider receives the question without anything that identifies the person asking. When the answer returns, the original details are restored before it is shown. In photos, faces are covered and location data is removed on the device, before they are sent.
On Ultra plans, messages can also be sealed on the device and opened only inside a secure server that Secure AI cannot see into, so the message is unreadable to Secure AI on its way to the AI provider.
Encryption and data storage
Conversations, documents and everything else you save are stored encrypted. A small amount of information, such as dates used for sorting, is stored unencrypted.
Secure AI holds the encryption key and shares it with your signed-in devices, which is what allows a conversation started on one device to continue on another. Because Secure AI holds the key, it is technically able to open this data. What that means in practice is set out below.
Data access
The list below shows what Secure AI is able to access, not what it does access. It is published because anything a company can open, it can be compelled to disclose.
AI providers and training
Nothing written in Secure AI is used to train AI models, by Secure AI or by any AI provider. This is a term of the agreement with every provider Secure AI uses, so there is no setting to change: training is always off.
Team accounts
On team plans, administrators can set company instructions, keep a list of terms that are never sent to an AI provider, limit who may join by email domain, and download a log of changes to the team. Administrators see usage counts and the team’s total of details removed. They never see the content of anyone’s conversations, or a breakdown of those details by person.
Reporting a security issue
To report a security issue in Secure AI on any device or on the website, email company@secureai.one with enough detail to reproduce it. The same address is published at /.well-known/security.txt.
Secure AI replies within 3 working days and shares a plan within 10. It will not take legal action against anyone who reports a genuine issue in good faith, stays within their own account and allows time for a fix. Reporters can be credited by name on request. Please do not access other people’s accounts or degrade the service for others.
Comparison
Secure AI compared with a typical AI account. The second column describes a common setup rather than any one company, since it varies by provider and plan.
| Question | Secure AI | Typical AI vendor |
|---|---|---|
| Does the AI provider know which company is asking? | No — identity is stripped before the request leaves | Yes — the account identifies the customer |
| Who can read a conversation later? | Stored encrypted. We hold the key and do not read it | The vendor, depending on its policies |
| Is our data used for training? | No — on every plan | Usually off on team tiers |
| How many vendor contracts for frontier models? | One | One per provider |
| Can we reclaim a seat when someone leaves? | Yes — per person | Usually yes |
| SAML / SSO? | Google, Apple or Microsoft sign-in. Ask us about SAML | Generally available on enterprise tiers |
Common questions
What can the AI provider actually see?
Only the request, after identifying information has been removed. They don’t see your company, your people, or the original values.
Can we verify the redaction rather than trust it?
On Ultra, yes. The redaction runs inside an AWS Nitro Enclave, and before anything is sent your browser checks a fingerprint of the running code, signed by Amazon’s hardware. The fingerprint is published on our Proof page.
What can Secure AI see?
Your conversations are stored encrypted. Secure AI holds the key so they sync across your devices, and staff open them only when their work requires it. On Ultra, sealed chats are unreadable to Secure AI.
Can an admin read what their team writes?
No. Personal chats and notes belong to the person who created them. Shared work belongs to the team.
What happens when somebody leaves?
Their seat is revoked. Work they shared with the team remains available to the team; personal work remains private.
Do you rotate the team key when somebody is removed?
Removing a member stops their access right away. Anything they had already opened stays with them, as with any shared document.
Do you support SAML or SCIM?
Sign-in uses Google, Apple or Microsoft, and seats are managed in Secure AI. Talk to us if you need SAML or SCIM.
Do you have SOC 2 or a DPA?
Our DPA is available at secureai.one/dpa. We do not currently have SOC 2.
Can we pay by invoice?
Yes. Invoice billing is available on request.
Questions about this page can be sent to company@secureai.one.