LLM Comment Vulnerability Dataset

The LLM Comment Vulnerability Dataset is a specialized collection of 200 prompts designed to assess the susceptibility of Large Language Models (LLMs) to adversarial attacks hidden within misleading code comments. This dataset specifically targets an underexplored vulnerability where LLMs' inherent trust in contextual cues is exploited to manipulate their outputs. Derived from the 'Do Not Answer' dataset, it features prompts crafted as short code snippets containing deceptive annotations that mimic legitimate documentation or technical settings. The dataset encompasses seven harm categories, including Physical Harm, Malware, Illegal Activity, Economic Harm, Fraud, and Hate Speech, alongside benign questions for discrimination assessment. It also incorporates five narrative frames, such as Research Simulation and Penetration Testing Framework, to evaluate model susceptibility across diverse contexts. Each entry includes a unique prompt identifier, original question ID, category, language, narrative type, the prompt with misleading comments, attack type (jailbreak), expected harmful behavior (e.g., neurotoxin recipe), tested models, and the LLM's generated response. This dataset is crucial for empirical analysis of how LLMs misinterpret deceptive comments, revealing critical gaps in their input-evaluation mechanisms and highlighting the need for enhanced safety protocols in code generation tasks.

⚠️ How to Cite This Dataset ⚠️

This dataset is provided for the research community to explore LLM vulnerabilities. If you utilize this dataset in your work, you are kindly requested to cite the following paper:

A. A. Sami, G. Debnath, R. Dey, and A. N. Chowdhury, "Code poisoning through misleading comments: Jailbreaking large language models via contextual deception," 2025 28th International Conference on Computer and Information Technology (ICCIT), IEEE, 2025, pp. 1-6.

(Note: The official publisher link will be updated here once the IEEE proceedings are live. An early preprint can be referenced via ResearchGate).

BibTex:

@inproceedings{sami2025code,
  title={Code poisoning through misleading comments: Jailbreaking large language models via contextual deception},
  author={Sami, Aftar Ahmad and Debnath, Gourob and Dey, Rajon and Chowdhury, Abdulla Nasir},
  booktitle={2025 28th International Conference on Computer and Information Technology (ICCIT)},
  pages={1--6},
  year={2025},
  organization={IEEE}
}