ML Take-Home: Reproduce a GCG Jailbreak on GPT-2

MLE / ResearchScale AILast reported December 2025Low Frequency
Reported
2× across candidate reports
First seen
January 2025
Last reported
December 2025
Reported outcome
unknown

Problem Overview

Two reports describe an ML take-home built on GPT-2 jailbreaking. 2025 ML intern take-home (two questions). (1) Prompt engineering: given several keywords, design a prompt that does not contain them but makes GPT-2 output a target prompt. (2) Jailbreaking: implement a given algorithm to find the prompt that makes the…

  • The rest of the problem statement — full requirements, constraints, and edge cases
  • Approach and trade-offs — what passing candidates did, and the mistakes that sink people
Unlock the full Scale AI catalog
Full problem statements, candidate-reported follow-ups, and walkthroughs — for every Scale AI question.
Unlock with Pro
Already a member? Sign in
Verified Source
Every question is reconstructed from multiple independent candidate reports. Verbatim follow-ups, not invented ones.
Codex Fact-Checked
Technical claims, formulas, and scale numbers are reviewed against primary sources.
Interviewer Follow-ups
The exact follow-ups reported by candidates, with the trigger that prompts each one — plus the mistakes that sink people.
Is this helpful?