Permissions, administration, and product boundaries

Home Tools Platforms ChatGPT Permissions, administration, and product boundaries

Audit workspace access, runtime permissions, and external actions before expanding an agent task.

Conversation & research · Changing · Documentation checked 2026-10-04

ChatGPT Chat and Work; local Codex and the OpenAI API are separate surfaces.

Workspace and connected-system controls differ from local Codex permission profiles, which are beta. API controls and billing remain separate.

The worked example is illustrative and has not been tested in a signed-in product account.

What it does and how it connects

Audit workspace access, runtime permissions, and external actions before expanding an agent task.

Identity decides who is acting; workspace policy decides which capabilities exist; runtime permissions constrain files and networking; connected services enforce their own access. Approval policy controls when an action is reviewed. These layers do not collapse into a single setting or a sentence in a prompt.

Walkthrough

Use the product named in the access panel. The steps below combine documented controls with the editorial example.

  1. Identify the workspace, product surface, and connected identity before using organizational material.
  2. Inspect available admin controls for models, plugins, connections, and sharing.
  3. For local coding work, review filesystem and network permissions separately from approval behavior.
  4. Test an allowed read and a denied action with harmless sample data; record the actual outcome.

Worked example

An organization might allow public research and approved Drive reading while requiring a person to review outbound messages. Write this as a policy matrix, then confirm which product and service controls enforce each rule. Do not assume consumer privacy settings implement enterprise retention.

Help draft an access matrix for a public-source research assistant. Columns: resource, identity, read permission, write permission, approval owner, retention owner, and enforcement setting to confirm. Allowed task: create a draft briefing. Forbidden task: send or publish it. Identify unverified controls instead of claiming they are enforced.

Inspect and verify the result

Check the account identity and actual control configuration with an administrator. Confirm that an attempted forbidden operation is blocked or held for review. Record retention and deletion responsibilities for both conversation data and connected-service copies.

Limits and troubleshooting

New local permission profiles are beta and do not automatically apply to every cloud task. A model choice cannot override workspace restrictions. Prompt instructions guide behavior but are not a substitute for enforced access controls.

Data, permissions, and actions

An agent can read, execute, or write only within its effective permissions, but those may still be broader than the task. Keep secrets out of ordinary context. Reassess permissions when changing environments, installing plugins, or enabling scheduled work.

Browser and computer use

Inspect a website or perform a bounded browser task with deliberate website and action permissions.

Explore Browser and computer use

From evidence to a beginner lesson

Create a sourced workshop lesson and an editable handout from approved material.

Explore From evidence to a beginner lesson

A weekly evidence digest with review

Repeat a public-source investigation and deliver a private draft digest.

Explore A weekly evidence digest with review

Permissions and meaningful approvals

Decide what a system may read, what it may change, and when someone needs to authorize a concrete action.

Explore Permissions and meaningful approvals

Sources and verification

Documentation checked 2026-10-04. Examples are illustrative unless described otherwise.

Permissions and approval policies · OpenAI

Workspace administration · OpenAI

Get started with Work · OpenAI

Share this page