Jailbreak Detection

Jailbreak detection studies how to recognize malicious or policy-evading inputs before a model complies, including attacks hidden in long conversations, other languages, images, or obfuscated text. Approaches range from compact classifiers using lexical and semantic features to runtime monitors that inspect hidden states, token entropy, refusal signals, or output logits; some track risk across turns or combine multiple detectors. The field is expanding from prompt-only checks toward multilingual, multimodal, and decoding-time detection, with attention to latency and deployment cost. A central challenge is maintaining high recall at very low false-positive rates when attacks and benign inputs differ from training data, especially under adaptive attacks and realistic conversation distributions.