Skip to main content

Trust

The controls in place today, and where each is written down

The controls that are in place today, each with the page that describes it. MITHUNAI holds no third-party security certification today. Security questionnaires are answered on request, directly, by the people who built the platform.

Last updated

Controls a review will find

Each row says whether the control is in place for every organisation or still in beta, and links to the page that describes it.

  1. Each organisation’s data is kept apart

    In place

    The organisation a request acts in comes from the verified session or key, never from the request. Every query and every vector lookup is filtered by it, a request without one is refused, and a resource in another organisation reads as not found.

    Read how it works: Each organisation’s data is kept apart

  2. Access is checked on every request

    In place

    Membership, role, key revocation and key expiry are re-read on each request, so removing access takes effect at once. Nobody can grant a role above their own, and an API key cannot manage API keys or people.

    Read how it works: Access is checked on every request

  3. Nothing acts without a person

    Beta

    In MITHUNAI Team, an AI employee’s action with an effect outside MITHUNAI waits for a person to approve, edit or reject it, and access, password, payment and publishing actions always do. Beta. Access on request.

    Read how it works: Nothing acts without a person

  4. An audit trail of what was decided

    In place

    Authorisation decisions, allowed and refused, are recorded, as are API keys, widget embeds, member changes and every use of a saved connection. If the record cannot be written, the request fails rather than going ahead unrecorded.

    Read how it works: An audit trail of what was decided

  5. Secrets are encrypted at rest

    In place

    A saved connection’s token is encrypted before it is stored, never shown again, and sent only to the one service it belongs to. API keys are stored only as hashes, and the Gateway keeps provider keys in an encrypted vault.

    Read how it works: Secrets are encrypted at rest

  6. Everything you connect is treated as untrusted

    In place

    Credentials found in your content are redacted before indexing. Retrieved passages are fenced off from instructions, and a fetch cannot reach a private, loopback or cloud metadata address.

    Read how it works: Everything you connect is treated as untrusted

  7. Support sees your organisation only if you let it

    In place

    Support access is off until an owner or administrator allows it. A session it then opens is read-only, excludes every conversation, expires on its own and writes each request into your own audit history.

    Read how it works: Support sees your organisation only if you let it

  8. No training on your content, no telemetry

    In place

    MITHUNAI trains no model on your content. A deployment makes no telemetry or analytics call, and talks only to the hosts it is configured with.

    Read how it works: No training on your content, no telemetry

Where it runs

  • Hosted by us

    The console, API and widget at app.mithunai.com, on a server in Germany. There is no choice of region yet.

  • Your own cloud account or data centre

    On the Sovereign Custom plan, billed annually and arranged with us. There is no self-service installer.

  • Your own models

    Every deployment needs a chat model and an embedding model; it calls them only with the credentials it is configured with. Running with no outbound calls at all is a conversation to have before you plan around it.

What a deployment consists of and what it calls

Not offered today

  • Third-party security certifications or attestations
  • A published data processing agreement or sub-processor list: ask us
  • Single sign-on with your own identity provider, and SCIM provisioning: planned
  • A customer-facing export of the audit trail: built, not yet deployed
  • Answers limited to what each reader can open in the source: built, cannot be turned on yet
  • Customer-managed encryption keys
  • A choice of data region, or a published uptime commitment

Security questionnaires, on request

Send your security questionnaire through the contact page and it is answered directly. Nothing on it is answered with a claim this page does not make.

How each control works, in more depth: Security.

Production Readiness

Grounded answers from your own knowledge, a chat workspace for your team and one model gateway, on one account.

A 30-minute walkthrough on your own use case: connect a source, ask it questions, check the citations and see what it does when the answer is not there. Or ask the public playground now, with no account.