TwinGate: Stateful Defense Against Decompositional Jailbreaks in Multi-Turn Conversations

arXiv·medium signal

Decompositional jailbreaks split a malicious objective across multiple benign-looking turns that individually pass safety filters. TwinGate uses asymmetric contrastive learning to maintain state across turns and detect when a conversation trajectory is reconstructing prohibited content, even in untraceable traffic where session linking is unavailable. A practical defense for any multi-turn agent or chatbot deployment where adversaries can probe over extended conversations.