Feature Request: User-Friendly Permission Center and Allowlist for Codex Desktop

[mike535]

Feature Request: User-Friendly Permission Center and Allowlist for Codex Desktop

Category: Security, Privacy & User Experience

Dear OpenAI Codex Team,

I would like to propose a centralized, user-friendly permission management interface for the Codex desktop application.

As a Windows user, I appreciate the potential of autonomous AI agents to improve productivity, automate development tasks, and assist with complex workflows. However, configuring and verifying their security boundaries currently requires significant technical knowledge.

My proposal is to introduce a graphical Permission Center with the following capabilities:

Directory allowlist: Users should be able to explicitly select which folders Codex can read, modify, or access. All other directories should be inaccessible by default.

Application allowlist: Users should be able to choose which applications Codex may launch, control, or interact with.

Granular tool permissions: Provide separate controls for terminal commands, web access, browser automation, desktop control, file operations, and external integrations.

Temporary permissions: Allow one-time or time-limited authorization, with automatic revocation after a task is completed.

Activity and security logs: Provide a clear history of accessed files, modified files, executed commands, network connections, and permission requests.

Real-time permission overview: Users should always be able to see exactly what Codex can access and which privileges are currently active.

Secure-by-default profiles: Offer simple presets such as Read Only, Project Only, Ask for Approval, and Full Access, with clear explanations of their security implications.

Most importantly, these restrictions should be technically enforced by the operating system or sandbox, rather than relying solely on instructions given to the AI model.

During my own testing of Codex on Windows, I found that workspace-write restrictions prevented writing outside the project, but still allowed reading a file located outside the authorized workspace without additional approval.

I understand that this may be expected behavior, but it highlights the need for clearer permission boundaries and better visibility for users.

The goal is simple: users should be able to benefit from powerful AI agents without sacrificing control over their computers, personal files, and privacy.

Thank you for considering this suggestion.