A cheap activation probe screens all traffic and escalates only suspicious exchanges to a stronger classifier that sees both sides of the conversation. This brings jailbreakA prompt crafted to make a model ignore its own guidelines — usually through roleplay, hypotheticals, or encoding rather than a direct request.Full definition → defense overhead down to roughly 1% compute, a usable pattern for guardrail design.
Terms in this piece · Glossary
jailbreak — A prompt crafted to make a model ignore its own guidelines — usually through roleplay, hypotheticals, or encoding rather than a direct request.
