Membrane: A Self-Evolving Contrastive Safety Memory for LLM Agent Defense

Abstract

Despite advances in safety alignment, large language models remain vulnerable to continuously evolving jailbreaks. Existing fine-tuned safety classifiers cannot adapt to these evolving attacks, while adaptive memory-based guardrails tend to over-refuse benign queries that resemble stored attacks. We propose Membrane, a self-evolving guardrail built on Contrastive Safety Memory (CSM): each cell pairs the conditions for blocking a harmful query with those for permitting a superficially similar ben

Categories

Framework mappings

Suggested from the entry's categories.

@misc{choi2026membrane,
  title = {{Membrane: A Self-Evolving Contrastive Safety Memory for LLM Agent Defense}},
  author = {Minseok Choi and Seungbin Yang and Dongjin Kim and Subin Kim and Jungmin Son and Yunseung Lee and Jaegul Choo and Youngjun Kwak},
  year = {2026},
  month = jun,
  eprint = {2606.05743},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.05743}
}