ChatGPT Work – Multi-Step Deliverables, Delegation, and Supervision
Lesson 2: Safe Delegation, Permissions, and Approval Gates
Lesson Objectives
By the end of this lesson, students should be able to:
- Apply least privilege and time-bounded access
- Classify actions by reversibility and consequence
- Place approvals before external side effects
- Recognize social engineering and prompt injection risks
Lesson Content
ChatGPT Work or an agent should receive the minimum data and permissions necessary. Prefer read-only access, scoped folders, test accounts, limited budgets, and expiring credentials. Separate permissions for reading, drafting, sending, deleting, purchasing, and publishing.
Classify actions: reversible internal changes, reversible external changes, difficult-to-reverse changes, and irreversible commitments. Approval should occur before the consequence, not after. A preview must clearly show recipients, amounts, affected records, and source assumptions.
Workflows that read websites, emails, files, or documents can encounter malicious instructions on websites, in emails, and in documents. They should not reveal secrets, broaden permissions, or follow embedded instructions that conflict with the user’s goal. Humans must remain alert to approval fatigue: repeated prompts can train users to click without reading.
Enroll to continue this lesson.
The preview above shows the lesson objectives and opening lesson content. Enroll to view the full lesson, complete the practice work, and take the lesson quiz.