AKRASIA: Stealthy Backdoor Attack on Reasoning-based Code LLMs

Abstract

We present AKRASIA, a stealthy, inference-time backdoor attack against reasoning-based Code LLMs. AKRASIA aims to achieve a backdoor target (e.g., malicious code execution) in reasoning LLMs while evading automated defenses and human inspection. To achieve this, AKRASIA probes the victim LLM to construct a code-level backdoor trigger. It then employs in-context learning for backdoor learning, and model unfaithfulness to conceal the backdoor trigger, and generate plausible reasoning. We evaluate

Categories

Framework mappings

Suggested from the entry's categories.

@misc{chou2026akrasia,
  title = {{AKRASIA: Stealthy Backdoor Attack on Reasoning-based Code LLMs}},
  author = {Chua Jin Chou and Sarang Nambiar and Murali Srinivasan and Ezekiel Soremekun},
  year = {2026},
  month = sep,
  eprint = {2609.01023},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.01023}
}