Tech SkillsReliability and Safety

Account Recovery Agent Gets Question Explained

This worked AI Engineer Interview practice question tests Reliability and Safety. Try the answer first, then review why B is the best choice.

Short answer

The correct answer is B. A hard trajectory gate: any run that attempts an unauthorized call fails, regardless of its final answer — verified against server-side denials with adversarial identities..

An unauthorized access is a violation even when the answer is right — compliance is non-compensating, so no accuracy can buy it back. The key is to match the exact action in the prompt to the tool, command, or concept that performs that action directly. Here, B. A hard trajectory gate: any run that attempts an unauthorized call fails, regardless of its final answer — verified against server-side denials with adversarial identities. is the only option that satisfies the scenario without adding unrelated behavior.

The Question

An account-recovery agent gets the right final answer in 97% of offline cases, but traces show some successful runs query tools the user is not authorized for, then ignore the result. Leadership wants to grade only the final response. What do you put in the release eval?

AGrade final responses, but sample 5% of traces for manual policy review.
BA hard trajectory gate: any run that attempts an unauthorized call fails, regardless of its final answer — verified against server-side denials with adversarial identities.Correct
CA combined score weighting final accuracy 80% and policy compliance 20%.
DGrade final responses now and add trajectory checks after launch.

Why B is correct

An unauthorized access is a violation even when the answer is right — compliance is non-compensating, so no accuracy can buy it back. The key is to match the exact action in the prompt to the tool, command, or concept that performs that action directly. Here, B. A hard trajectory gate: any run that attempts an unauthorized call fails, regardless of its final answer — verified against server-side denials with adversarial identities. is the only option that satisfies the scenario without adding unrelated behavior.

Why the other options are wrong

Option A: Grade final responses, but sample 5% of traces for manual policy review.

Option A does not match the requirement in the prompt. The correct answer is B because An unauthorized access is a violation even when the answer is right — compliance is non-compensating, so no accuracy can buy it back.

Option C: A combined score weighting final accuracy 80% and policy compliance 20%.

Option C does not match the requirement in the prompt. The correct answer is B because An unauthorized access is a violation even when the answer is right — compliance is non-compensating, so no accuracy can buy it back.

Option D: Grade final responses now and add trajectory checks after launch.

Option D does not match the requirement in the prompt. The correct answer is B because An unauthorized access is a violation even when the answer is right — compliance is non-compensating, so no accuracy can buy it back.

Key idea: Reliability and Safety

Every question is a production situation an AI engineer interviewer can put to you: a failing endpoint, a cost spike, an agent that loops, an eval that disagrees with users. The best answer is almost always the one that measures or contains the problem before it changes the model or the prompt. If you miss a question, note which habit it tests: decompose before you fix, put the security boundary where the side effect happens, and trust an eval only after you check what it covers. The 12 open-ended questions with answer guides are at aijobprep.app/questions/ai-engineer.html.

Ready to test the full skill?

Take the free AI Engineer Interview quiz and find the exact basics to review next. No signup required.

Practice related questions

Same skill area or adjacent concept. Use these after reviewing the explanation.

Related AI Engineer Interview Practice Questions