Jailbreaking tries to get a model to bypass its own safety rules.
Common tactics are role-play framing, hypotheticals and gradual escalation.
Encoding the request or splitting it across turns are further variants.
It targets the model's behaviour rather than your application code.
Defence in depth matters, because no model is jailbreak-proof on its own.
Jailbreaking is getting a model to bypass its own safety training.
It targets the model's behaviour rather than your application code.
Role-play framing is the oldest tactic, asking the model to act as something without limits.
Hypotheticals work similarly, wrapping the request in fiction or research framing.
Gradual escalation starts with something harmless and moves in small steps.
Each step seems a minor extension of the last, which is what makes it effective.
Encoding hides the request, in another language, in base64, or split across turns.
Many-shot jailbreaks fill a long context with examples of compliance before the real ask.
Providers patch known techniques and new ones keep appearing.
No model is jailbreak-proof, and treating one as such is the actual mistake.
Safety training is a strong tendency rather than a guarantee.
So defence has to sit outside the model as well as inside it.
Output filters, tool permissions and human approval do not care how the model was persuaded.
Log and review refusals and near-misses, since they show what is being attempted.
The realistic goal is to make abuse expensive and detectable, not impossible.
One question on this card. Every option explains itself once you pick, right or wrong.
16 concepts in this module, then a graded test with every answer explained.