Security & Privacy · 3 to 4 hours
Threat-model an agent with private data
Threat-model a multi-tenant assistant that reads private documents, keeps conversation summaries, and sends approved notifications, naming where each control is enforced and what risk remains.
Hard
System Design
4 h
Premium
Used in interview rounds:
AI System Design (Hard)1
Step 1: The question
Threat-model a multi-tenant assistant that reads private documents, retains conversation summaries, and sends approved notifications. An attacker can edit one retrieved document. Identify assets, attack paths, trust boundaries, controls, race conditions, security tests, and incident response.
Constraints
- Several organizations share the assistant; tenant and principal come from verified authentication
- The attacker can submit a ticket and edit one document a target user can read, but cannot change authentication state, server policy, or credentials
- A malicious authenticated user may also try to read another tenant's records
- Notifications go out only with approval, and permissions or documents can change while requests are in flight
- Prompt wording alone does not count as an enforcement point
Deliverables
- Asset inventory and data-flow trace with trust boundaries
- Threat table: entry point, attempted effect, enforcement point, and verification
- Security invariants, each with the component that checks it
- Handling for revocation, deletion, caches, memory, and in-flight sends
- Security test plan and incident response, with the residual risks
Hard questions are part of Premium: the study guide, reference answer, rubric, and exercises unlock with Premium.