StealthBox AI

Control is the architecture, not a setting.

StealthBox AI treats control as a property of the system, not an add-on. Eleven control domains govern the local AI estate — each readable as an executive outcome or a technical mechanism — so data, decisions and authority stay with the organisation. Air-gapped or isolated deployment can be designed where required and technically validated. No certification or zero-risk claim is implied.

Data boundaries

Where your data lives — and where it stops.

A StealthBox estate is built around trust zones. The customer perimeter is the outer boundary; everything inside it is governed by your identity, your permissions and your audit. Connectivity outward is explicit, never assumed.

Customer perimeterYour network. Your data. Your rules.
AI runtimeInference inside the perimeter.
Knowledge & modelsBounded, versioned, revocable.
Agents & toolsScoped actions, human approvals.
IntegrationsExplicit, audited, least privilege.
Controlled connectivityOffline-capable where required.

Air-gapped or isolated deployment can be designed where required and technically validated — not assumed for every site.

The control framework

Eleven domains. One governed estate.

Each domain is a control you can point to — for a minister, an auditor, a CISO or an engineer. The executive line states the value; the technical line states the mechanism.

01

Data control

Where enterprise data resides and who controls it — the system sees only what you authorise and retains only what your policy allows.

TechnicalAccess-scoped corpora, retention and deletion policies, data residency rules, no training on customer data without explicit approval.

02

Local processing

Workloads can execute inside customer-controlled infrastructure — data does not leave your perimeter to be processed by a third party.

TechnicalOn-prem and private-hosted inference, local model serving, in-room compute for sensitive workloads, no egress for inference.

03

Network boundaries

Where the system can reach is a policy decision, not a default — connected, controlled-connected, private or isolated architectures.

TechnicalEgress controls, service boundaries, segmentation, deployment topology as configuration, connectivity class selectable per environment.

04

Identity & access

Users, roles, permissions and service identities follow least privilege — the AI operates under the same identity boundaries as your people.

TechnicalFederated identity, RBAC and ABAC, scoped tool permissions, least-privilege agent authorisation, service identity boundaries.

05

Knowledge control

Which sources the AI can read, which it cannot, and where each answer came from — knowledge is bounded and attributable.

TechnicalSource permissions per corpus, access boundaries on retrieval, retrieval and source provenance, access-scoped indexing.

06

Model control

Which models are approved, which versions are live, how they are routed and how they change — models are governed, not automatic.

TechnicalApproved model catalogue, version pinning, model gateway and routing, change control, rollback, local weight management where supported.

07

Agent & tool control

Which actions AI can perform, which require deterministic rules, and which require a human approval — boundaries set by policy.

TechnicalTool permission scoping, allowed/forbidden action lists, deterministic rules for predictable steps, human-in-the-loop for consequential actions.

08

Human authority

AI can prepare, analyse and recommend. Authorised people remain responsible for consequential decisions.

TechnicalHuman-in-the-loop gates, configurable autonomy levels, escalation paths, approval workflows, explicit decision rights by action class.

09

Auditability

Requests, retrieved knowledge, actions, decisions, approvals and system events should be traceable where the selected architecture supports it.

TechnicalImmutable audit logs, retrieval provenance, action provenance, exportable evidence trails, decision lineage from prompt to outcome.

10

Monitoring & evaluation

Operational monitoring, evaluation, error review, quality review and controlled updates — the system is observed, not assumed.

TechnicalRuntime telemetry, evaluation harnesses, error and drift detection, quality review cycles, controlled model and knowledge updates.

11

Business continuity

Backup, recovery, deployment support and lifecycle management according to the selected architecture — resilience designed in, not hoped for.

TechnicalBackup and recovery procedures, staged deployment, rollback, lifecycle support, model and knowledge versioning for restoration.

Principle

Security by architecture, not by promise.

We do not sell certifications we cannot show you, nor claim air-gap support we cannot deliver. No system is zero-risk. Where a control depends on the environment — offline operation, sovereign cloud, national residency — we say so and design to it, sized to what your mission actually requires.

What stays yours
  • —Your data, your perimeter, your retention rules.
  • —Your identity boundaries, applied to the AI.
  • —Your approvals before high-stakes actions.
  • —Your audit trail, exportable and complete.
  • —Your models and knowledge, versioned and revocable.
Next

Sovereign by design — shipped by method.

The control framework is the what. The deployment method is the how.