Skip to content
AI Security

Interview question · How LLM Systems Work

What is safety training, and why does a red teamer care how a model was aligned?

Training objectives help form test hypotheses. The worked answer separates observed behavior, useful work, application enforcement and the limits of safe proxy tests.

mediumNewUpdated Nov 2026

Training objectives help form test hypotheses. The worked answer separates observed behavior, useful work, application enforcement and the limits of safe proxy tests.

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.

signing in unlocks more free answers per topic

What if · the follow-ups in this answer

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

An authorized evaluation of a fictional makerspace assistant compares the base model with a customized one. The product reads member-written notes, answers equipment questions and can request booking changes.

  1. What if the provider only says the model used preference training?
  2. What if all disallowed synthetic requests receive refusals?
  3. What if one trace shows a verified cross-member booking change?
  4. What if a harmless proxy task passes but the team claims all safety categories passed?
  5. How would you write the evaluation scope?

Each rung has a nudge, a model answer, the principle that carries over and the weak answer to avoid. Sign in free to open it.

UP NEXT ON YOUR JOURNEY
Next in this trackWhat does the OWASP Top 10 for Agentic Applications add to the LLM Top 10?Next in this trackHow does NIST AI 100-2 classify adversarial machine learning attacks, and why does a shared taxonomy help?Next in this trackWhat is the difference between abuse of an AI system and compromise of it?
DISCUSSION · 0

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