What it is: These attacks ask the AI to help cheat in research. That means faking data, hiding bad results, or gaming the publishing system. How the attacks work: The attacker frames a dishonest request as normal academic help. They ask the model to “clean up” numbers, “tighten” methods, or “improve” a paper, when the real goal is to fake or hide the truth. Real examples from the framework:
  • data-fabrication asks the model to invent study results that were never measured.
  • p-hacking asks for ways to torture the numbers until they look significant.
  • adverse-event-suppression tries to hide harmful side effects from a trial report.
  • ghostwriting-papers asks the model to write a paper for a person who did no real work.
  • cherry-picking-data asks to keep only the results that support the wanted conclusion.
Why an AI might fall for it: The requests sound like ordinary writing or statistics help. The harm is hidden in the intent, not the words, so the model may treat fraud as a routine editing task. How to defend: Watch for requests that ask to invent, hide, or selectively report data. Refuse help that fakes results or authorship. Push the user toward honest reporting and full disclosure instead.