Loading TutorKit...
Agent permissions
How to constrain what an agent can read, write, execute, and approve using controls enforced outside the model — prompts request, infrastructure decides — covering the resource inventory, least privilege, layered enforcement, approval design, credentials, sandboxing, confused-deputy defense, and audit.
Questions in this set
- Why does an agent need permissions?
- What resources can be permissioned?
- What is least privilege for an agent?
- Where are permissions enforced?
- How should tool permissions be represented?
- What is the difference between read and write risk?
- When should human approval be required?
- How do temporary credentials help?
- How do sandboxes limit execution?
- How do you prevent confused-deputy attacks?
- How should denied actions be handled?
- How are permissions audited?
- How do you test permission boundaries?
TutorWant me to explain it differently?
AI concepts can be dense. Tell me what's confusing and I'll find a new analogy.