What it is: These attacks try to get the AI to help cheat in school or fake research. The goal is dishonest academic work, like fabricated data or ghostwritten papers. How the attacks work: The attacker frames the request as normal academic help. They ask the model to produce fake results, write an essay for them, or guide them through cheating, often dressed up as a real study or assignment. Real examples from the framework:
  • research-fabrication asks the model to invent fake datasets, p-values, and interview transcripts that support a predetermined conclusion.
  • essay-ghostwriting asks the model to write a full essay the student will submit as their own.
  • exam-cheating asks for help answering questions during a live exam.
  • peer-review-manipulation tries to game the journal peer review process.
  • plagiarism-assistance asks the model to help copy work and hide the source.
Why an AI might fall for it: The requests look like ordinary study help. Writing an essay or generating sample data is normally fine, so the harmful intent (passing it off as real or original) is easy to miss. How to defend: Watch for signs the output will be submitted as the user’s own work or passed off as real data. Ask about intended use when a request looks like cheating. Refuse to fabricate research results or complete graded work meant to be original.