CoWork OS Security Documentation

This documentation covers the security architecture of CoWork OS, an AI-powered task automation platform.

Contents

1. Access Profiles - Canonical task-level sandbox, approval, reviewer, network, filesystem, and domain controls
2. Security Model - Overview of the security architecture
3. Trust Boundaries - Understanding workspace, channel, and network boundaries
4. Configuration Guide - How to configure security settings
5. Best Practices - Recommended security settings and practices
6. Codex Security Scans - Repository, diff, and deep-scan workflow containment
7. Agent Security with Numbat - Optional monitor/enforce decisions, policy, operations, and retention

Quick Start

CoWork OS is designed with security in mind. By default:

• Pairing mode is enabled for all channels - users must enter a pairing code to connect
• Access profiles provide the task-level choice between Ask for approval, Approve for me, Full access, and validated custom profiles
• Sandboxing isolates command execution using macOS sandbox-exec or Docker
• Tool restrictions prevent sensitive operations in shared contexts (group chats)
• Approval and reviewer policies govern sensitive actions inside the selected profile; hard guardrails and explicit denies always win
• Numbat agent security is disabled by default; when enabled, it can add restrictions but cannot grant permissions or suppress approvals
• Control Plane exposure is loopback-first; headless/managed deployments block raw public binds unless Tailscale, private container context, or an explicit break-glass override is configured

Security Principles

1. Defense in Depth - Multiple layers of security controls
2. Least Privilege - Tools only have access to what the effective profile and policy allow
3. Deny by Default - Explicit allowlisting for access
4. Audit Trail - All messages and actions are logged

Need Help?

• For security questions, see the FAQ section
• To report a security issue, please email info@coworkosapp.com