OpenAI introduced the Admin plugin for ChatGPT Work and Codex on August 25, 2026. It gives authorized workspace administrators a conversational interface for reviewing activity and credit usage, managing members and groups, checking permissions, changing access, and handling usage-limit or spending requests.

What the Admin plugin changes

Instead of moving between analytics, settings and reports, an admin can ask a question, inspect the result, request a supported change and confirm its outcome in one conversation. OpenAI says the plugin can also automate recurring checks and route pending usage requests to Slack or Microsoft Teams for an authorized decision.

  • Usage visibility: review ChatGPT Work and Codex activity, credit consumption and approaching limits.
  • Identity administration: add or remove members and update groups for onboarding, offboarding and team changes.
  • Access controls: inspect effective permissions, diagnose access issues and control features or models by role or group.
  • Budget workflows: review, approve or deny usage and spending requests with current usage as context.

The plugin does not grant extra privileges. OpenAI says every request is mapped to a supported read or write action under the user’s existing role, workspace policy and approval requirements. Broader-impact changes can be reviewed before application, and the result identifies whether the action completed and what changed.

Why this matters for QA engineers

This turns AI administration into another agentic workflow that needs functional, security and audit testing. Natural-language convenience must not weaken least privilege or make a sensitive change harder to trace.

  • Test the same request with admin, delegated-admin and ordinary-member roles; verify both the response and the actual workspace state.
  • Use ambiguous prompts to confirm the plugin asks for clarification before changing groups, limits or model access.
  • Validate approval gates for high-impact changes and prove that unauthorized Slack or Teams reviewers cannot approve requests.
  • Compare before-and-after permissions, budgets and memberships against the plugin’s structured confirmation.
  • Exercise partial failures, retries and duplicate requests so an automation cannot apply the same write twice.
  • Retain prompt, actor, approval, action and outcome data as auditable test evidence.

A practical QA scenario

Create a synthetic group near its credit limit and submit a request to raise that limit. Verify the plugin reports the current usage, routes the request only to approved reviewers, blocks an unauthorized decision, applies one authorized change exactly once, and returns a confirmation that matches the Admin Console. Repeat with a denied request and a reviewer timeout.

The release is aimed at workspace administration, but the broader lesson is relevant to every QA team adopting agents: conversational control planes need the same deterministic authorization, idempotency and evidence standards as conventional admin APIs.

Source