AIMentor — AI Expertise Learning Platform

Untimed. Answers are graded when you submit, with explanations and spaced-repetition scheduling.

Question 1easy
What is permission propagation for an agent?
The agent runs with the app's broad service credentialsThe agent acts as the user, inheriting that user's permissions and no moreThe agent has no permissionsPermissions are set by the model at runtime
Question 2medium
Where should authorization checks for an agent's tools live?
In the system prompt, trusting the model to filterIn the tool/data layer, enforced with the acting user's identityNowhere — the model handles itOnly in the UI
Question 3hard
A tool queries the DB with admin credentials and the agent decides which rows to return. Why is this dangerous?
It's slowerA prompt injection can make the agent return anyone's rows — the model chooses data it shouldn't have authority overIt uses more tokensIt only works offline
Question 4medium
What is the MOST effective protection for PII in an LLM pipeline?
Sending it at higher temperatureMinimization — don't transmit PII the task doesn't need in the first placeEncrypting the response onlyUsing a bigger model
Question 5medium
How does PII redaction let a model reason without receiving raw personal data?
It deletes the data permanentlyPII is replaced with placeholders before the call and restored in the output, so the model only sees placeholdersIt fine-tunes the model on the PIIIt skips the model entirely
Question 6hard
What must you check before sending data to a model provider? (Select all that apply)
Whether the provider trains on your inputsData retention periodData residency / where it's processed, and compliance postureThe provider's marketing slogan
Select all that apply.
Question 7medium
For the most sensitive/regulated data, which deployment keeps data fully under your control?
A consumer chatbot tierA self-hosted open-weight model, so data never leaves your infrastructureAny public API with a long retention windowEmailing it to the provider
Question 8hard
What is the RAG-specific multi-tenant footgun?
Embeddings are too slowA shared vector index retrieved without a tenant filter can pull another tenant's chunk into the prompt — a silent breachChunks are too smallThe model forgets the tenant
Question 9hard
Where must the tenant scope be enforced for retrieval and DB reads?
In the prompt, asking the model to only use tenant A's dataIn the data layer — every retrieval/read carries the tenant filter, never left to the modelIn the UI onlyIt doesn't need enforcing
Question 10medium
For high-sensitivity tenants, what is stronger than a per-query filter?
A longer promptPhysical separation — a per-tenant index/namespace/database so an app bug can't leak across boundariesA bigger modelHigher temperature
Question 11medium
Where should API keys and DB credentials live in an agent system?
In the system prompt so the agent can use themIn a secrets manager / env vars — never in prompts, code, or traces (which log prompts)Hardcoded in the clientIn the vector index
Question 12medium
What scope field must filter every retrieval and DB read in a multi-customer system to prevent cross-customer leaks? (two words)
Question 13hard
What principle (two words) means the agent inherits the acting user's rights and no more?

0 / 13 answered

Access & Data Protection — Practice quiz