TwinGate: Stateful Defense against Decompositional Jailbreaks in Untraceable Traffic via Asymmetric Contrastive Learning

Abstract

Decompositional jailbreaks pose a critical threat to large language models (LLMs) by allowing adversaries to fragment a malicious objective into a sequence of individually benign queries that collectively reconstruct prohibited content. In real-world deployments, LLMs face a continuous, untraceable stream of fully anonymized and arbitrarily interleaved requests, infiltrated by covertly distributed adversarial queries. Under this rigorous threat model, state-of-the-art defensive strategies exhibi

Categories

Framework mappings

Suggested from the entry's categories.

@misc{sun2026twingate,
  title = {{TwinGate: Stateful Defense against Decompositional Jailbreaks in Untraceable Traffic via Asymmetric Contrastive Learning}},
  author = {Bowen Sun and Chaozhuo Li and Yaodong Yang and Yiwei Wang and Chaowei Xiao},
  year = {2026},
  month = apr,
  eprint = {2604.27861},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.27861}
}