Brilliant AI
Get started
Security

Built so that nothing bad can happen quietly .

The platform doesn't ask you to trust the model. It uses pre-determined workflows, strict gates and an audit trail you can read.

Eight pillars

What makes AI safe enough to actually plug into a real business.

  • Per-tenant isolation

    Every customer's data lives in its own logical store. Cross-tenant reads are structurally impossible, not just discouraged.

  • Agent data gate

    Every read passes through a per-agent gate. Customer-facing agents can never read owner-only fields.

  • Audit trail

    Every action — read, write, decision — is recorded with the agent, the slug, the disposition and the reason.

  • Human-in-the-loop by default

    Customer-facing writes default to draft-for-review. Owner sets the autonomy threshold per channel + per action.

  • Reversible by design

    Drafts before sends. Notes before publishes. Approval queues for anything money- or reputation-touching.

  • UK / EU data residency

    Customer data stays on UK and EU infrastructure. AI models run in the EU. No off-shore data transfers.

  • Least-privilege writes

    Each writable field is owned by exactly one agent. An unknown or out-of-scope write is refused, not silently dropped.

  • Fabrication guardrail

    The platform refuses to invent numbers, prices or facts. If it doesn't know, it says so.

What we won't do with your data

Your business data is yours. Full stop.

  • We never train the underlying AI models on your business data. Your customer conversations, your records, your numbers - none of it becomes training data for a foundation model. The models we use are either off-the-shelf and frozen, or fine-tuned on permitted datasets only.
  • We never sell your data to a third party. Not aggregated, not anonymised, not "in a partnership". The only third parties who ever see your data are the infrastructure providers we list in our subprocessor list, and only the minimum necessary to deliver your service.
  • We never use cross-tenant inference. Even when we learn patterns that help every tenant (eg "this kind of message is usually a complaint"), the learning happens on anonymised aggregates with k-anonymity >= 5 and differential privacy noise, never on a single tenant's identifiable data.
  • You can export or erase on request. UK GDPR Articles 15 (access), 17 (erasure) and 20 (portability) honoured in code - not in policy. The security.txt contact is the fastest route.

Compliance posture

  • UK GDPR + EU GDPR aware Lawful basis recorded per channel; data subject rights honoured; retention policies enforced in code.
  • Article 9 special-category protection Sensitive personal data carries an explicit consent grant + a dedicated key + a mandatory audit trail.
  • Encryption at rest + in transit Per-tenant keys; TLS 1.3 across the wire; never plain-text secrets in logs.
  • Daily backups + point-in-time recovery 14-day PITR on the database; 35-day PITR on the key-value tables; rolling daily VM snapshots.
  • Vulnerability + dependency scanning Workflow security scans every workflow; CVEs surfaced daily; signed container artifacts.

Responsible disclosure

Found a vulnerability? Email security@brilliantstudios.co.uk. We respond within one business day, fix critical issues within seven, and credit researchers who'd like to be named.