AIMentor — AI Expertise Learning Platform

13 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
What are input and output guardrails?
Two different LLMsChecks run before the model (block/sanitize input) and after it (block/fix unsafe or leaky output)Embedding filters onlyA type of vector index
Question 2easy
What is prompt injection?
Injecting SQL into a databaseUntrusted text containing instructions the model follows, overriding your intentAdding more tokens to a promptA way to speed up inference
Question 3medium
What is INDIRECT prompt injection?
The user types 'ignore your instructions'The attack hides in content the model reads (a web page, PDF, email, or RAG chunk), not in the user's messageInjecting a prompt into the model weightsA slow-motion denial of service
Question 4medium
Why is prompt injection so hard to stop at the root?
Models are too smallYour instructions and untrusted data share one channel (the context window), and the model can't reliably tell them apartIt only affects open-source modelsBecause temperature is too high
Question 5hard
Why is 'add a line telling the model to ignore malicious instructions' NOT a real fix for injection?
It uses too many tokensThat line is itself just more text an attacker can out-argue; it's not a security boundaryIt only works on GPT-4It actually is a complete fix
Question 6hard
Which are REAL structural defenses against prompt injection? (Select all that apply)
Least privilege on tools (minimum capabilities)Human-in-the-loop approval for consequential actionsValidating the proposed tool call/action against policy before executingTrusting the model to police itself with a strong system prompt
Select all that apply.
Question 7hard
What is the correct mental model for defending against injection?
Make the model impossible to trickAssume the model CAN be hijacked, and ensure a hijacked model still can't do damageOnly accept input from logged-in usersDisable all tools permanently
Question 8medium
How does least-privilege tool design defend against injection?
It makes the model fasterAn agent that lacks a destructive capability can't be tricked into using it — a read-only agent can't deleteIt encrypts the promptIt hides the system prompt
Question 9medium
For content safety/moderation, why screen BOTH input and output?
To double the costA safe input can still produce unsafe output, so the response must be screened tooOutput moderation is optionalInput moderation is unnecessary
Question 10medium
Which are managed content-moderation options?
NumPy, PandasOpenAI Moderation, Azure AI Content Safety, Llama GuardDocker, Kubernetespgvector, Qdrant
Question 11hard
What role should an injection-detection classifier play in your defenses?
The single, complete defenseOne layer of defense-in-depth atop structural controls (isolation, least privilege, human gates) — it misses novel attacksA replacement for least privilegeUnnecessary if you have a good prompt
Question 12medium
What is the #1 LLM security risk (two words) where untrusted text carries instructions the model obeys?
Question 13hard
What security principle (two words) — giving an agent only the minimum tools/permissions it needs — limits injection damage?

0 / 13 answered

Guardrails & Content Safety — Final Exam