TwinGate: Stateful Defense Against Decomposed Jailbreaks in Untraceable Traffic via Asymmetric Contrastive Learning

TwinGate is a stateful dual-encoder defense framework that uses asymmetric contrastive learning to identify decomposed jailbreaks in untraceable anonymous request streams. It clusters malicious fragments that are intent-consistent but semantically different into a shared latent space, and uses a parallel frozen encoder to suppress false positives. On a dataset of more than 3.62 million instructions and 8,600 malicious intents, it achieves high malicious-intent recall and extremely low false positives under a strict causal protocol, with throughput and latency better than stateful and stateless baselines.