Security you can check, not take on trust.
These are the controls in the SphereIQ product today, each described as it is built. Where something is not built yet, it is marked Not yet rather than left out.
Who can get in
Sign-in goes through your identity provider, and each person gets only the role they are given.
- Single sign-onGoogle Workspace — optionally limited to your company’s domain — or any SAML 2.0 identity provider.
- SCIM 2.0 provisioningUsers are created, updated and deactivated from your identity provider, so someone removed there loses access here.
- RolesAdministrator, user and viewer. Every administrative request checks the role on the server — hiding a button in the interface is not the control.
- Scoped API tokensProgrammatic access uses bearer tokens that each carry an explicit set of scopes, rather than a person’s full rights.
- Sign-in limitsFive attempts per account in fifteen minutes. Passwords, for accounts that do not use single sign-on, are stored as salted PBKDF2 hashes.
- Not yetBuilt-in multi-factor sign-inWith single sign-on, your identity provider’s MFA applies. MFA for password accounts is not released.
What each person can see
Answers are built from the documents the person asking is allowed to read.
- Permission-aware retrievalEach knowledge source carries its own access rules, applied in the retrieval query itself — before any passage reaches a model.
- CitationsAnswers cite the document and passage they came from, so a reader can check the source rather than trust the summary.
- Not yetFolder-level grantsAccess is granted per source today. Granting one folder inside a source is not released.
What the AI takes in and gives out
Checks run in the request path on every chat message, not as a report afterwards.
- Prompt-injection detectionDirect attempts, and indirect ones — instructions hidden in a document or a tool result — are detected before they reach the model.
- Secret scanningKeys, tokens and credentials in a message are caught before it is sent to a model.
- Personal-data maskingSocial security numbers, card numbers, email addresses and phone numbers are detected and masked by content policy. Detection is pattern-based.
What is recorded
A record you can hand to an auditor, and check without asking us.
- Audit logActions are written to an audit log with who acted, what they did and when. The application only ever adds to it.
- Signed eDiscovery exportAdministrators can export records as newline-delimited JSON. The manifest carries a SHA-256 of the records and an Ed25519 signature, so anyone can confirm offline that nothing was changed.
- Signed receiptsReceipts are Ed25519 signatures over canonical JSON, and the public key is published — so they can be verified without us.
- Not yetStreaming to your SIEMForwarding the audit log live to Splunk, Datadog or a CEF collector is not released.
How credentials and traffic are protected
Credentials for your systems are sealed, and traffic is encrypted in transit.
- Sealed credentialsTokens and keys for connected systems are sealed with AES-256-GCM, with a separate key derived for each purpose, so a value sealed for one use cannot be opened as another.
- Encrypted in transitHTTPS only, with HSTS, and a content security policy that limits what the application may load and call.
Where it runs, and what it calls
Choose the deployment that matches your data boundary.
- Three ways to run itCloud, hosted by Sphere; Dedicated — a single-tenant environment in your own AWS account, managed by Sphere; or Private — in your own cloud, where you hold the key that seals stored credentials.
- AI providersModel calls go from your deployment to the providers you enable — OpenAI and Anthropic today — through their APIs. Cohere reranking is used only if you turn it on.
- No model trainingSphereIQ does not train models on your data.
Certifications
Stated as they stand.
- SOC 2 program in progressThere is no SOC 2 report yet. This page will say so when there is one.
- For your own compliance workComply keeps a register of your AI systems against the EU AI Act, with 40 policy templates including GDPR and the EU AI Act.
Questions procurement asks
Answered from the product, including where the answer is “not yet”.
Do you have a SOC 2 report?
Not yet. The SOC 2 program is in progress. Until there is a report, this page says so rather than implying one.
Can we use our own identity provider and its MFA?
Yes. Sign-in works with any SAML 2.0 identity provider or with Google Workspace, and your identity provider’s multi-factor rules apply. Users can be provisioned and removed with SCIM 2.0.
Can an answer draw on a document the person asking cannot open?
Each source’s access rules are applied in the retrieval query, before anything reaches the model, so answers are built from the passages that person is allowed to read.
Can we run it in our own cloud?
Yes. Private runs in your own cloud, and you hold the key that seals stored credentials; disk and database encryption are your infrastructure’s. Dedicated runs as a single-tenant environment in your AWS account, managed by Sphere.
Can we take a verifiable record out?
Yes. Administrators can export records as signed newline-delimited JSON. The manifest carries a SHA-256 of the records and an Ed25519 signature that can be checked offline with the published public key.
Which AI providers see our data?
Only the ones you enable: OpenAI and Anthropic today, called through their APIs from your deployment, and Cohere for reranking if you turn it on. SphereIQ does not train models on your data.
Send us your security questionnaire.
We answer it from the product as it is built, and say plainly where something is not built yet.