// Security · Privacy

Security, governance and data protection

What your security team needs to know before authorizing a proof of concept.

Data Omnis · Context Layer for Enterprise AI

Nothing enters your network.

The runtime operates in outbound-only mode: connections always go out from Data Omnis to the sources you authorized. There is no open port, no reverse tunnel and no installed agent that accepts inbound connections. Your network's exposure surface does not grow when you adopt the platform.

Each client, its own world.

  • Organization and workspace: no resource is shared between clients.
  • Database isolation: enforced by Row-Level Security in PostgreSQL: queries are filtered by the tenant identifier at the database level, not just in the application.
  • Authentication: identity managed by Keycloak, with a signed token per request and administrator and member roles.
  • Connection credentials: encrypted at rest with a dedicated key. They never appear in a log, on screen or in an API response.
  • Language model: you can use your own provider key. The key belongs to the organization and is encrypted; no user sees or carries it.

Guaranteed by structure

  • Sensitive columns do not enter the query. Fields marked as personal data are excluded in the layer that builds the query. The model does not receive them, and therefore cannot reveal them.
  • Aggregates only in dashboards. Count, sum, average and grouping. There is no path to read an individual row this way.
  • Writing requires human approval. The agent proposes; an authorized person approves, and there is a record of who approved.
  • Department scope. The content of an area's dashboard is restricted to that area's assets, checked at execution.

Being hardened, and we say so openly

In free conversation, the agent's reach is determined by the configured context package and by the sources connected to the workspace. We are evolving this control so that the per-department restriction is also verified structurally in conversational reads. In environments with strict segregation between areas, today we recommend one workspace per area, which makes the isolation structural on every path.

We would rather state this than present a guarantee we could not stand behind in an audit.

No access without a record.

Every data read (by a person or by an agent) generates a record in an append-only trail, which the application can neither edit nor delete:

FieldWhat it is for
WhoThe user or agent that originated the access
WhatThe resource queried and the dataset involved
WhenDate and time
PurposeThe stated purpose, linked to the legal basis
Policy decisionAllowed or denied, and by which rule
Treatment appliedWhat was masked or excluded

Denied accesses are recorded too. In an audit, that is as relevant as the allowed ones: it shows the policy was active and was enforced.

Personal data protection under Brazilian law

RequirementHow we meet it
Purpose (art. 6, I)Recorded on every access, not inferred afterwards
Necessity and minimizationPersonal columns excluded by default; work with aggregates
Security (art. 46)Per-tenant isolation in the database, encrypted credentials, encrypted traffic
Accountability (art. 6, X)Append-only trail, exportable for the data protection officer
External transferSending by email blocked by default; only to authorized domains, with a record
Data subject rightsThe records make it possible to locate in which operations a piece of data was used

Advertising cookies

If you accept in the cookie notice, the site loads the Meta pixel and the LinkedIn tag. They tell us whether a visit came from an ad and whether it ended in a contact request, so we know which campaigns work. If you decline or do not respond, nothing from Meta or LinkedIn is loaded. To change your choice, clear this site's data in your browser.

Frequently asked questions

Is client data used to train models?

No. When you use your own provider key, your contractual terms with that provider apply.

What is stored?

Metadata (names of tables, columns, descriptions and relationships) and the aggregated results that make up dashboards and reports.

What if an agent tries to access something outside what is allowed?

The access is denied and the event is recorded with the reason, visible in the governance dashboard.

Who can change the audit trail?

Nobody through the application. It is append-only by design.

What is the minimum permission you ask for?

As little as possible. In cloud file connectors, read access to the bucket you specify. In OneDrive, the drive of the user who authorized it. Access to the organization's files is an explicit opt-in, and that does involve your administrator.

You saw it work with an example. Run it with your data.

We reply within 1 business day.

I could not send it right now. Try again or reach us on WhatsApp.

Received.Thank you, our team will be in touch soon.

Start with what your security team needs to see.

During the proof of concept, every access is recorded and available for your team to consult, in real time, inside the platform itself.

Context Diagnostic: a conversation with your security team about the access records and, if it makes sense, a day with your operation.

Talk to an AI expert →
Contact us