jailbreak — A prompt crafted to make a model ignore its own guidelines — usually through roleplay, hypotheticals, or encoding rather than a direct request.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
Why it matters
Pays up to $15,000 for universal jailbreaks and gives selected researchers access to an unreleased safety mitigation system before public deployment, with universality defined as bypassing guardrailsThe checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.Full definition → across a set of specific harmful questions.
