Trust · The Doctrine

Security

Last Updated: 10 July 2026

Never Trust The Client.

One rule runs through every part of LLM Vault: whatever a sender says about its own payload is a claim, not a fact. The server re-classifies and re-checks everything on arrival — every payload, every source, every time. A client stamping its own request “Public” changes nothing, because the platform decides for itself.

Fails Closed

Real Data Risk

If a gate cannot establish that data is safe to proceed, the call does not proceed. Safety is never assumed, and there is no override that skips the question.

Fails Open

Optimisation Features

If a convenience feature is unavailable, your work continues without it. Convenience never holds traffic hostage — and it never becomes a back door around the safety gates.

That asymmetry is deliberate, and it is the second half of the doctrine.

The Controls, By Layer

Pre-Model Gating

Classification, DPIA triage, residency and data-risk checks all complete before any model receives the request — governance runs on the request path itself.

Encryption & Keys

Encrypted in transit and at rest. Your BYOK provider keys are encrypted at rest, and credentials are stripped on ingest — unrecoverable because they were never written.

Isolation & Roles

Each organisation's data, policies and audit trail are scoped to its own workspace by construction — isolation is how the system is built, not a filter applied afterwards. Inside it, access follows Owner, Admin and Analyst tiers.

Tamper-Evident Evidence

Every governed call mints a provenance record with SHA-256 hashes and an audit log entry. Alter the record and the hashes no longer agree.

What each step of the request path records is itemised, field by field, on the Data Usage page; how little survives retention is on Privacy.

How Do We Relate To Compliance Frameworks?

Honestly. LLM Vault is engineered with the expectations of GDPR, the EU AI Act, NIST AI RMF, ISO 27001, ISO/IEC 42001, SOC 2, NIS2, the CIS Controls, AESCSF and PSPF in view — the control and evidence trail those frameworks expect an organisation to be able to produce is precisely what the platform generates.

But alignment is not certification: using LLM Vault does not make your organisation certified against anything, and we claim no formal certifications during beta. It is tooling and evidence to help you manage AI risk — the assessment itself remains yours, your auditor’s, or your regulator’s.

Report A Vulnerability

Found something? We want to know. Email john@cybereq.io with what you found and, if you can, the steps to reproduce it. We respond to responsible disclosures, and we credit the researchers who make them.