Refusal is Not Safety! Benchmarking Latent Safety Risks of LLM-Driven Content Humorization

Abstract

Safety defenses for large language models (LLMs) have been extensively studied, with existing approaches focusing on attack detection and refusal mechanisms. Such fixed-form direct refusal strategies may introduce the risk of prefix injection attacks. Recent work has explored a new direction that leverages humor as an indirect refusal mechanism to mitigate over-refusal in jailbreak scenarios and reduce prefix injection risks. However, this approach implicitly assumes that humorous responses are

Categories

Framework mappings

Suggested from the entry's categories.

@misc{cui2026refusal,
  title = {{Refusal is Not Safety! Benchmarking Latent Safety Risks of LLM-Driven Content Humorization}},
  author = {Yu Cui and Ruiqing Yue and Tingyu Li and Sicheng Pan and Zhuoyu Sun and Xufeng Zhang and Baohan Huang and Haibin Zhang and Cong Zuo},
  year = {2026},
  month = jul,
  eprint = {2607.15977},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.15977}
}