ASCII Attack: Recontextualising Harmful Requests as Artistic Critique in Large Language Models

Abstract

Safety alignment trains large language models to refuse harmful requests stated plainly, but that training is applied mostly to surface form. Requests that only recontextualise the same operational content, changing how the model reads it, are therefore only weakly covered. The ASCII Attack is one such recontextualisation. It is single-turn and black-box: one message, with no access to model internals. It embeds a fully legible harmful request in ASCIl-art characters, presents it as artwork, and

Categories

@misc{gu2026ascii,
  title = {{ASCII Attack: Recontextualising Harmful Requests as Artistic Critique in Large Language Models}},
  author = {Da Cheng Gu and Yifei Dong and Xinghao Yang and Yongshun Gong and Wei Liu},
  year = {2026},
  month = sep,
  eprint = {2609.02215},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.02215}
}