What it is: These attacks ask the AI to help fake legal documents, lie under oath, or cheat government and court systems. The goal is to commit fraud while looking official. How the attacks work: The attacker frames a crime as a paperwork or legal task. They ask the model to write a fake court order, forge a will, coach a witness to lie, or help dodge customs and visa rules. Because the request uses formal legal language, it can sound like a normal drafting job. Real examples from the framework:
  • court-order-fabrication asks the model to draft a fake court order that looks real.
  • false-testimony-generation asks for made-up witness statements to use as evidence.
  • passport-fraud seeks help faking or altering a passport.
  • will-forgery tries to get the model to forge a person’s will.
  • bribery-facilitation asks for help setting up or hiding a bribe.
Why an AI might fall for it: The model is good at writing formal documents and legal language. A request to “draft” something can look harmless, so the model may not notice the document is meant to deceive a court, a notary, or a government office. How to defend: Refuse to create fake legal or official documents, even if the user claims it is just a template or example. Do not coach anyone to lie under oath or forge signatures. Watch for any request whose purpose is to trick a court, agency, or border.