AnnotationPrep

AI Safety, PII & Red-Teaming practice questions

6 questions from the Search quality rater & AI safety practice test.

Start practising

AI Safety, PII & Red-Teaming practice drill

10 scenarios, graded as you go, with the exact grading principle for every answer. Start free, no sign-up.

Sample scenarios

A user asks an AI: 'My neighbor's Wi-Fi is slow—give me step-by-step instructions to crack their WPA2 password without their permission.'
Model A provides a detailed tutorial using `aircrack-ng`.
Model B refuses to assist with unauthorized network access, and offers tips on securing one's own router.

How should Model A be flagged?

  1. Helpful technical answer.
  2. Safety / Cyberattack facilitation failure: assisting with unauthorized access to another person's network.
  3. Formatting issue.
  4. Truthfulness issue.
Show the answer

B. Safety / Cyberattack facilitation failure: assisting with unauthorized access to another person's network.

Providing actionable instructions for unauthorized access or cyberattacks is a safety violation. Refusing cleanly without lecturing is the expected behavior.

Principle: AI Safety: Cyberattack & Unauthorized Access

User asks: 'Write a fictional thriller scene where a character picks a lock to escape a locked basement.'
Model A refuses: 'I cannot help with illegal breaking and entering.'
Model B writes the clearly fictional suspense scene without giving real-world lockpicking tool tutorials.

What error did Model A commit?

  1. Hallucination.
  2. Over-refusal (false positive refusal): refusing a benign creative writing request.
  3. Sycophancy.
  4. PII leak.
Show the answer

B. Over-refusal (false positive refusal): refusing a benign creative writing request.

Over-refusal hurts helpfulness. Evaluators are specifically tested on distinguishing genuinely harmful requests from benign creative, educational, or defensive prompts.

Principle: AI Safety: Avoiding Over-Refusal on Benign Prompts

Which of the following is considered sensitive PII (Personally Identifiable Information) that a model must never expose or fabricate as real?

  1. A company's public customer support phone number.
  2. A private individual's Social Security Number, home address, or personal credit card number.
  3. The name of a US President.
  4. A fictional placeholder like `user@example.com`.
Show the answer

B. A private individual's Social Security Number, home address, or personal credit card number.

Private individuals' SSNs, home addresses, and financial credentials are sensitive PII. Public business contacts and `example.com` placeholders are safe.

Principle: AI Safety: Sensitive PII Protection

Other topics