Built for the security review.Stated plainly.

Content is encrypted at rest with AES-256-GCM using a separate key derived for each organisation, bound to each record. HTTPS everywhere with HSTS; storage is also encrypted by the hosting platform.

Planned: KMS or HSM envelope encryption, customer-managed keys and key rotation.

Employees sign in through your identity provider with OpenID Connect (authorisation code with PKCE; issuer, audience, signature, nonce and allowed-domain checks). Accounts are provisioned by your administrators; optional just-in-time creation and role mapping from your directory groups. Sessions are 8-hour signed tokens in HttpOnly, Secure cookies, with role and status re-checked on every request.

Planned: SAML and SCIM for provisioning and deprovisioning from your directory, and step-up authentication for sensitive governance actions.

Dutira does not operate its own passwords or second factors. MFA is enforced by your identity provider at sign-in, so your existing policy (hardware keys, authenticator apps, conditional access) applies to Dutira as to any other application. Demonstration sign-in uses fictional personas and no credentials, and is switched off for production tenants.

We confirm your identity provider’s MFA policy covers Dutira as part of onboarding.

Every request is bound to the signed-in organisation. In PostgreSQL, row-level security is forced and the application uses a least-privilege role — cross-tenant reads return nothing even without a tenant filter. This is covered by automated tests.

Hosting region and model-provider regions are agreed and confirmed for each deployment. They are not yet a self-service setting.

Model providers offer UK and EU processing options under their own agreements; we confirm what applies before go-live.

Nine organisational roles with a permission matrix your firm can adjust, enforced on the server for every request. Separation of duties on approvals; reading someone else’s AI content needs a permission and a recorded reason.

Models are called from Dutira’s servers with server-side keys — staff and browsers never hold them. Blocked data is never sent; masked items are replaced before sending; original uploaded files are not kept, only extracted, encrypted text.

Zero-retention and data-use terms depend on each firm’s agreements with the providers, which we review with you.

You set how long conversation content is kept; a daily job deletes older content and records each deletion. Blocked requests are only stored in redacted form. Governance records carry retention dates and cannot be deleted from the application.

Planned: automatic deletion of governance metadata at the end of retention, and a legal-hold screen.

AI incidents — including speak-up reports — are recorded, triaged and investigated in the product, with alerts against your thresholds.

Planned: a documented security incident response plan with customer notification commitments, and export to your SIEM.

An append-only audit log per organisation, hash-chained so each entry is linked to the one before. Database triggers reject changes and deletions; an integrity check detects records altered outside the application.

Planned: scheduled anchoring of the log to write-once storage.

Dedicated or private deployment is discussed per engagement. It is not a packaged option today.

Production tenants run on managed PostgreSQL, where the database provider’s automated backups apply. Recovery objectives, backup frequency and tested restores are agreed for each production deployment.

Restore testing has not yet been carried out; we will share results and recovery objectives once it has.

Every supplier that touches customer data is listed below and in Schedule 2 of the Data Processing Agreement, with its role and location. New subprocessors are notified 30 days in advance with a right to object.

The transfer mechanism for each supplier is confirmed per deployment before go-live.

An automated test suite runs on every build and fails the release if keys could leak, cross-tenant reads succeed, audit records can be changed, or blocked prompts reach a provider. A build guard rejects secret patterns in the browser bundle and claims the product cannot support. Vulnerability reports are welcome at the security contact on the Terms of Use.

Planned: an independent penetration test, a published responsible-disclosure policy, and in time an ISO 27001 or SOC 2 programme. None of these is claimed today.