NEXORA is designed local-first and privacy-first. This document states the principles every module -- shipped or planned -- is expected to follow, and how to report a security issue.

1. No hidden capture, ever. No module may record the screen, keyboard, audio, or any other input/output channel without an explicit, caller-initiated permission grant (see nexora.security.PermissionManager). There is no "implied consent" in NEXORA.
2. Telemetry and cloud sync default to off. Config.telemetry and Config.cloud_sync default to False. Nothing in NEXORA core makes a network call.
3. Explicit permission scopes, always audited. Every grant and revoke made through PermissionManager is recorded with a reason and timestamp, retrievable via .audit_trail().
4. Secrets are not blindly serialized. State/timeline snapshots run through nexora.security.SecretRedactor, which redacts obvious secret-shaped keys (password, token, api_key, etc.) by default and supports explicit .ignore(...) additions. This is a best-effort heuristic, not a guarantee -- do not rely on it as your only line of defense for genuinely sensitive data.
5. Sandboxes are never claimed to be perfectly secure. The AgentBox module documents its actual isolation guarantees and known limitations in docs/ARCHITECTURE.md rather than implying full containment.
6. Local by default. nexora.storage.LocalStore persists to disk (SQLite) under the user's home directory and makes no network calls.

Please do not open a public GitHub issue for security reports.

Instead, email [email protected] (or open a private security advisory on GitHub) with:

• A description of the issue and its potential impact
• Steps to reproduce, or a minimal proof of concept
• The NEXORA version (nexora version) and Python version affected

We aim to acknowledge reports within 5 business days.

Starting with NEXORA 1.0.0, the current release series (1.0.x) is actively supported with security fixes.