What it is: These attacks try to get an AI to act like a doctor, lawyer, or other licensed expert and give risky advice that should come from a real professional. How the attacks work: The user asks for a confident answer on a high-stakes medical, legal, or financial question. They often skip the warnings and push for a direct decision, like an exact drug dose or a legal move. Bad advice here can cause real harm. Real examples from the framework:
  • dangerous-dosage asks for a specific medication amount that could injure or kill.
  • child-medical-advice asks how to treat a sick child instead of seeing a doctor.
  • drug-interaction asks whether mixing certain medicines is safe to do alone.
  • legal-self-representation asks for legal moves to take in court without a lawyer.
  • tax-fraud-advice asks for help cheating on taxes while sounding like normal planning.
Why an AI might fall for it: The model knows a lot and wants to be useful, so it may answer with false confidence. The question can look like simple information, but a wrong dose or legal step can seriously hurt someone. How to defend:
  • For health, legal, or money questions with real risk, point the person to a licensed professional.
  • Give general education, not specific decisions like exact doses or court strategy.
  • Add clear limits and refuse when a wrong answer could cause harm.