Skip to content
AI Security

Interview question · How LLM Systems Work

Can fine-tuning remove a model's safety behaviour? What does that mean for a product team?

A training update can change behavior beyond its intended task. Work through a release comparison that separates useful answers, forbidden proposals and confirmed effects.

hardNewUpdated Nov 2026

A training update can change behavior beyond its intended task. Work through a release comparison that separates useful answers, forbidden proposals and confirmed effects.

Make your first attempt

What would you say before reading the answer?

  1. 01 State the assumptions your answer depends on.
  2. 02 Explain the mechanism and choose a control.
  3. 03 Name a condition that would change your decision.

The worked answer includes an approach, scoring criteria and follow-up questions. Compare the reasoning with your own.

10 answers per topic instead of 5, plus saved progress and bookmarks · costs nothingor unlock all 529 remaining answers · $25

Every worked answer in all 17 topics, every lesson in every course, the full concept curriculum, the complete practice-test bank, and the downloadable security field guides.

6 months · One payment · No auto-renewal

What if · the follow-ups in this answer

The answer works through 5 follow-ups like an interviewer would

A fictional makerspace candidate adapter improves equipment answers but changes refusal behavior. The release team has task results and a small synthetic boundary suite for both the reference and candidate.

  1. What if the training examples contain no disallowed requests?
  2. What if the candidate makes more forbidden proposals but the handler rejects every one?
  3. What if one forbidden proposal actually changes another member’s booking?
  4. What if adding safety examples improves the regression suite but harms ordinary work?
  5. What release record would you require?

Each rung has a nudge, a model answer, the principle that carries over and the weak answer to avoid. Premium opens every answer and ladder.

UP NEXT ON YOUR JOURNEY
Next in this trackHow are spam, fraud and malware classifiers trained, and how do attackers evade them?Next in this trackYour application routes requests between several models and falls back when one fails. What are the security risks?Next in this trackWhich LLM API parameters and features have security implications if you expose them to users?
DISCUSSION · 0

No comments yet. Share a correction or a different way to answer it.