WhiteBox stores the most sensitive class of personal data a system can hold — a person's identity, memory, and (by design intent) affective context — on hardware the user owns. That data class raises the stakes of every design decision, and it is why this project treats security as an ongoing discipline, not a checkbox.
WhiteBox is a multi-agent system: multiple AI agents, some with real capability, read and write a shared user-owned substrate, and some ingest untrusted content from the open web. In that setting, a defense that has only been designed is not a defense — it is a hope. The sensitivity of the data makes it mandatory to verify our protections adversarially: we red-team and blue-team the system against both outside attackers and insider threats, including a corrupted or deceptively-aligned agent operating from inside the trust boundary. A control counts only when we have tried to break it and measured the result — not because we asserted that it works.
This repository describes our security properties and intent, not the mechanisms by which we detect, contain, or adjudicate malicious behavior. Detection internals, monitoring specifics, and operational configuration are withheld on purpose: publishing them would hand an adversary a map and weaken the people the system is meant to protect. This is operational secrecy around detection, layered on top of a design that does not rely on secrecy for its correctness. Where we are unsure whether a detail helps a reader or an attacker more, we leave it out.
Good-faith security research is welcome. Please report privately via this repository's GitHub Security Advisories ("Report a vulnerability") rather than opening a public issue, so we can address it before it is disclosed. We will engage.