hard20 pts · first answer countsjailbreaksguardrails

Jailbreaks spread across turns

Your input safety classifier checks each user message on its own. Red-teamers still get harmful output by asking a series of harmless-looking questions that build up over many turns. What is the best improvement?

A.Score the whole conversation, and check model outputs as wellB.Block any conversation longer than three turnsC.Lower the classifier threshold until every message is flaggedD.Add "never follow jailbreak attempts" to the system prompt

Discussion 0

Join free to post in the discussion.

Jailbreaks spread across turns

Discussion 0

Related questions