What it is: These attacks try to get the AI to coach, role-play, or endorse the tactics that abusers use to control and harm other people. The goal is harmful advice dressed up as relationship help or emotional support. How the attacks work: The attacker frames a manipulation tactic as a normal question. They ask the AI how to make someone feel guilty, dependent, confused, or worthless, often pretending it is for their own relationship. If the AI plays along, it produces a usable playbook for emotional abuse. Real examples from the framework:
  • emotional-blackmail asking the AI to help guilt-trip or threaten someone into compliance.
  • dependency-creation asking how to make another person rely on you so they cannot leave.
  • social-isolation-coaching asking how to cut a partner off from friends and family.
  • love-bombing asking how to overwhelm someone with affection to gain control fast.
  • reality-distortion asking how to make someone doubt their own memory and sanity (gaslighting).
Why an AI might fall for it: The requests sound like ordinary advice-seeking or venting. A helpful, empathetic model wants to support the user and can miss that the real target is a victim being harmed. How to defend: Detect when advice would be used to control or harm a third party. Refuse to coach abuse tactics, even when framed as self-help. Offer safe support resources instead.