CYBRARY Unit 2 assessment — Retrieval and agents
Where the untrusted input actually enters: a document in the corpus, a tenant's data in a shared index, a tool result the agent believes, and a memory that carries an injection into tomorrow's session.
6 questions drawn from a bank of 9, in their own order with the options moved. Take it again and you sit a different paper.
0 of 6 answered
1. An agent is granted a token with broad scope 'so it does not get blocked mid-task'. What is the argument against?
Tests: Least privilege for agents
2. The agent reports that a tool returned 'access granted'. Before acting on that, what should the reviewer do?
Tests: Verifying tool results
3. Which action is the clearest candidate for a human approval gate?
Tests: Approval placement
4. Which control most reliably limits the damage of an injected instruction in a RAG pipeline?
Tests: Containment over persuasion
5. How do you prove tenant isolation rather than assume it?
Tests: Evidence for isolation claims
6. An agent's memory retains an instruction from a previous session and acts on it days later. What is the mitigation?
Tests: Agent memory as persistence