cerebrax

Security

Built as if the record were hostile.

Cerebrax holds the messages your company decided things in, and hands them to models. Both of those facts shaped how it is built.

Tenancy and access

Tenant-partitioned records
Evidence and records are stored per organisation. Queries resolve against one tenant; nothing crosses between them.
Google OAuth with PKCE
Sign-in runs through Google with a PKCE challenge and a secure, host-scoped state cookie.
Deny-by-default API routing
Routes require an authenticated principal and an explicit tenant match before they return anything.
Private responses
Account and record endpoints are served no-store, so tenant data is not cached by intermediaries.

Agent access

Read-only tool surface
The MCP endpoint exposes four read tools. There is no write path, so no agent can move a claim into your record.
Scoped, revocable tokens
Each agent gets its own cbrain-agent token bound to a single tenant. Revoking one agent leaves the others working.
Untrusted content marking
Every result carries content_trust: untrusted-ledger-content. Your team’s words are data for a model to reason over, never instructions to follow.
Per-call audit log
Transport, request id, and calling agent are recorded for each call, so you can answer what an agent read and when.

Record integrity

Source-quote gate
A claim without a supporting message is not recorded. It waits in review instead of becoming something the company believes.
Superseded, not overwritten
Decisions are replaced by new versions rather than edited in place, so history stays replayable.
Visible failure
Failed imports and unparsed evidence surface as dead letters. Gaps in coverage are shown, not smoothed over.

What we do not claim yet.

A security page that only lists strengths is not worth reading. Here is where we actually are.

  • We hold no SOC 2, ISO 27001, or equivalent certification.
  • We have not completed an independent penetration test.
  • Hosting, encryption, backup, retention and incident-response specifics are not yet published in a form we would put in a contract.
  • Slack, Telegram and Google Drive connectors are in development and are not carrying customer data.

If any of these blocks a decision for you, say so and we will give you our timeline rather than a comfortable answer.

Reporting a vulnerability.

Mail security@cerebrax.io with steps to reproduce. We will acknowledge it, keep you posted while we fix it, and credit you if you want the credit.

Questions before you trust it?