Why it matters
Jailbreak detection shapes robustness. Understanding shapes defense.
Advertisement
The architecture
Known pattern matching.
Classifier trained on jailbreak dataset.
LLM-as-judge on prompt intent.
Advertisement
How it works end to end
Pattern: DAN, base64, role-play cues.
Classifier: BERT-based on known jailbreaks.
LLM judge: analyze intent.
Combine signals.