Why it matters

Jailbreaks bypass safety. Defense is arms race; layered helps.

Advertisement

The architecture

Safety RLHF training.

Refusal classifier.

Output filter for policy.

Jailbreak defenseSafety RLHFtraining-timeRefusal classifierruntimeOutput filterpolicy checkOWASP LLM01/LLM06; layered defense mandatory; monitor for new patterns
Jailbreak defense.
Advertisement

How it works end to end

Safety training.

Refusal classifier.

Output policy filter.