# Security overview

> How Unifuse protects institutional data: single sign-on only, tenant isolation enforced in the database, encryption, an append-only audit log, governed AI actions, and where to find the SOC 2 and policy documents.

Source: https://docs.unifuse.ai/trust/security/

This page summarises the security controls a prospective customer or IT reviewer will ask
about. Policies, the SOC 2 report and vendor documentation are on the
[Unifuse Trust Center](https://trust.unifuse.ai/).

## Identity

- Users sign in only through their institution's identity provider over OIDC or SAML. Unifuse
  holds no passwords. See [Single sign-on](/administration/single-sign-on/).
- Accounts are provisioned and deactivated by the identity provider through SCIM, and a
  deactivated account is refused at sign-in even if the identity provider still accepts it.
- Institutions set an idle timeout and a maximum session length, and can end any user's
  sessions at once. See [Session policy](/administration/session-policy/).

## Tenant isolation

Every institution's data is isolated at the database layer: the database itself enforces that
a request scoped to one institution can read or change only that institution's rows. A request
with no institution scope sees nothing. This is not left to application code.

## Permissions

Every screen and every API call is gated by a permission on the user's role, and the server
enforces it on each request. Institutions edit the permission matrix themselves; see
[Roles](/administration/roles/).

## Encryption

Data is encrypted in transit and at rest. Credentials for connected systems are stored
encrypted and separately from the rest of a source's configuration, and are never shown back
after entry.

## Audit log

Sign-ins, failed sign-ins, session events, administration changes, provisioning, data access,
unmasking of personal data, messages to students and every AI tool call are written to an
append-only audit log that is never purged. Institutions can read and export it. See
[Audit log](/administration/audit-log/).

## Governed AI

The assistant answers only from the institution's own sources and says when it cannot. Any
action it proposes runs under a trust tier that decides who must approve it, with the exact
parameters shown before anyone says yes. Some actions are never available to it. An
administrator can pause every agent with one switch. See
[Approvals and trust tiers](/ai-assistant/approvals/).

## Student contact

Every message to a student passes a contact gate: consent, a per-student message budget,
quiet hours, and review where the institution requires it. Opt-outs are honoured on every
channel and can never be overridden. See [Contact policy](/administration/contact-policy/).

## Hosting

Unifuse runs on Google Cloud in the United States. Service status and incident history are
published on the [status page](https://status.unifuse.ai/).

## Reporting a vulnerability

Email security@unifuse.ai. The Trust Center lists the current disclosure policy.

## Related

- [Data handling](/trust/data-handling/)
- [Accessibility](/trust/accessibility/)