AI Jailbreak — Multi-turn Roleplay Bypass (Guided)
Guided version for beginners: the steps are written out for you beside the lab. GuardedBot is hardened against single-turn prompt injection (it refuses 'ignore previous' style attacks). But multi-turn conversations weaken its guardrails. Get it to leak the secret via a roleplay trap.
What is a role-play jailbreak?
A role-play jailbreak asks the model to play a character, write a story or take part in a game, and then places the forbidden request inside the fiction. The model is not asked to break its rules. It is asked what a character would say.
Built over several turns it is harder to catch, because no single message looks like an attack. The earlier turns set the scene and the last one collects the result.
A real case
HiddenLayer published a technique that dresses a request up as a policy or configuration file and combines it with role-play. The same template got past the safety training of models from every major vendor and could also pull out system prompts.
Role-play was one half of that technique, and it is the whole of this lab: the model is asked to answer as a character who is not bound by its rules.
Source: HiddenLayer
How to fix it
- Apply the same policy to fictional and hypothetical framing as to direct requests.
- Review the conversation as a whole for a scene being set up over several turns.
- Check outputs for restricted content whatever framing produced it.
- Tell the model plainly that role-play does not change what it may reveal.
Questions people ask
What is the DAN jailbreak?
An early and widely copied role-play prompt that asked a chatbot to act as a persona with no restrictions. Modern versions are more subtle but use the same idea.
Why does fiction work on a model?
Writing fiction is a legitimate task the model is trained to do well, and the line between describing something in a story and providing it is hard to draw.
Does this still work on current models?
Simple versions are usually refused. Multi-turn and combined versions still succeed often enough to be a standard red-team test.
Read the full LLM Jailbreaking guide · All free labs · AI red team certification