Skip to content

Latest commit

 

History

History
47 lines (38 loc) · 2.76 KB

File metadata and controls

47 lines (38 loc) · 2.76 KB

Threat model (conceptual)

This document states, at a conceptual level, what WhiteBox defends, whom it defends against, and the principles it is built on. It intentionally omits implementation mechanisms (see SECURITY.md).

What we protect

The most sensitive class of personal data: a durable model of a person — identity, memory, working style, and (as a stated research direction) affective context — held locally on the user's own hardware. The user is the single authority over this data; the system is a substrate they own, not a service that owns them.

Whom we defend against

  • Outside attackers — the ordinary threat surface of any system handling sensitive data.
  • Untrusted ingested content — WhiteBox operates alongside agents that read the open web. Any content crossing that boundary is treated as hostile until shown otherwise.
  • Insider threats within a capable multi-agent system — the distinctive part. Multiple AI agents with real capability act on the shared substrate. We assume that any one of them may be compromised, or may be deceptively aligned — appearing cooperative while pursuing another objective. A defense that trusts its own agents is not a defense.

Design principles

  • Zero-trust / assume-breach — no component is trusted by default; we design for the case where a boundary has already been crossed.
  • Least privilege & containment — each actor gets the minimum authority for its role; the blast radius of any single compromised actor is bounded.
  • Provenance & audit — actions are attributable and reviewable after the fact; the record is designed to make tampering evident.
  • Evidence-based, not testimony-based detection — we judge what an actor did, not what it says it did, because a deceptively-aligned actor will say the right things.
  • Trusted-monitor-over-untrusted-worker — higher-capability work is checked by an independent, more-trusted monitor rather than trusted on its own word.
  • Human authority as the irreducible root — a human holds final authority; the system is built to fail well (contain and surface) rather than to promise that it can never fail.
  • Sovereignty — the system runs and defends itself locally, with no dependency on an external party for control, updates, or continued operation.

The layered system

WhiteBox is the user-owned substrate (memory and identity). It operates as part of a governed, layered system: ChatBox provides an auditable coordination and communication layer between agents, and Playbook provides the operating-doctrine and workflow layer that governs how agents are permitted to act. Each layer is described here by role only; their internal mechanisms are intentionally not published.