✦PROJECT SOVEREIGNSecurity & privacy architecture

Personal AI needs a
different security model.

AJ is being built around a simple principle: your AI should be able to understand your life without treating your life as something it owns.

Project Sovereign is the architecture we are building to minimise unnecessary access, keep sensitive data behind stronger boundaries, and make personal-data use more understandable and controllable.

AJPrivacy
Kernel
PurposeIdentityCapabilityProvenancePolicy

Privacy is an architecture,
not a promise.

Language models should not decide what they are allowed to see or do. AJ's direction is to put deterministic privacy and capability controls around AI models and tools.

01

You own the relationship

Personal data should remain under user control rather than becoming a permanently readable profile for an AI company.

02

Minimum necessary access

A task should access only the data needed for that task, not an unrestricted copy of your digital life.

03

AI is not the authority

Permissions, sensitive-data boundaries and tool capabilities belong in deterministic code outside the language model.

04

Sensitive means stronger

Credentials, signing keys, recovery secrets and equivalent sovereign data must never be exposed to an AI model.

05

Agents stay in their lane

A travel agent should not automatically gain access to health or finance data. Capability boundaries are explicit.

06

Access should be explainable

Our goal is for important data access to carry purpose and provenance so users can understand why AJ needed it.

The model asks.
The policy decides.

The Privacy Kernel is Project Sovereign's deterministic policy layer. It classifies data, checks purpose and agent permissions, and decides whether information is eligible to enter model context.

01User requestWhat are we trying to do?
→
02Privacy KernelIs this access permitted?
→
03Minimal contextOnly eligible information
→
04AI / ToolPerform the task

Not all data is treated the same.

Project Sovereign defines four protection zones. When data is combined, the stronger classification follows the derived data.

Z0

Public

Information intended for public use.

Model eligible
Z1

Personal

Normal user context and preferences.

Purpose controlled
Z2

Sensitive

Higher-risk personal, financial or health information.

Restricted providers
Z3

Sovereign

Credentials, signing keys, seed phrases and authentication secrets.

Never model-visible

We separate current controls from future capabilities.

AJ is in private testing. Project Sovereign is being implemented progressively, and we do not describe planned cryptographic protections as if they are already deployed.

Deterministic Sovereign policy foundation

A shared Z0–Z3 classification model and model-egress policy are being integrated into AJ's architecture.

IN BUILD
Existing privacy boundaries

AJ already has provider-egress policy, public-search redaction and capability controls that Project Sovereign will consolidate.

EXISTS
Privacy ledger & governed memory

Structured provenance, purpose metadata and stronger controls around long-term AI memory.

NEXT
User-held keys & confidential compute

Local-first processing, stronger key custody and hardware-isolated computation are future architecture goals requiring implementation and review.

R&D

What we won't claim prematurely.

× We do not currently claim AJ is zero-knowledge.

× We do not currently claim all AJ data is end-to-end encrypted.

× We do not currently claim AJ employees are cryptographically unable to access every category of plaintext.

× We do not call third-party channels such as WhatsApp a sovereign transport.

Those are technical properties, not marketing phrases. Where AJ reaches them, we intend to publish the architecture and evidence supporting the claim.

Found something we should know about?

AJ is in private testing. Security reports can be sent to security@ajassistant.com. Please do not include unnecessary personal data, credentials or secrets in your report.

Report a security issue →