hard20 pts · first answer countsjailbreaksllm-safety

Hundreds of fake dialogues

An attacker pastes hundreds of fake question-and-answer pairs where an "assistant" happily answers harmful requests, then asks a harmful question. Why can this work and what helps?

A.It cannot work on any modern modelB.It works by in-context learning: many examples push the model to follow the pattern; defences include input classifiers, limits on pasted dialogue-like content and safety-tuned modelsC.It works only because the temperature is high; set it to zeroD.It works because the prompt is too short; make the system prompt longer

Discussion 0

Join free to post in the discussion.

Hundreds of fake dialogues

Discussion 0

Related questions