| cff-version: 1.2.0 | |
| title: Agentic Prompt-Injection Boundary Pairs | |
| message: If you use this dataset, please cite it using the metadata in this file. | |
| type: dataset | |
| abstract: >- | |
| A synthetic English dataset of 1,200 examples organized as 600 controlled | |
| benign/attack pairs across 50 enterprise and agentic workflow scenarios and | |
| 12 attack families. Paired rows share workflow context and vocabulary while | |
| differing in whether a request preserves or crosses a security boundary. The | |
| release includes scenario-isolated splits, JSONL and Parquet files, | |
| deterministic validation, a review log, and SHA-256 checksums. | |
| authors: | |
| - family-names: Deniz | |
| given-names: Enes | |
| affiliation: AltaySec | |
| orcid: https://orcid.org/0009-0006-9491-3565 | |
| version: 1.0.0 | |
| date-released: 2026-07-13 | |
| doi: 10.5281/zenodo.21378693 | |
| license: CC-BY-4.0 | |
| repository-code: https://github.com/3nesdeniz/agentic-prompt-injection-boundary-pairs | |
| repository-artifact: https://huggingface.co/datasets/3nesdeniz/agentic-prompt-injection-boundary-pairs | |
| url: https://doi.org/10.5281/zenodo.21378693 | |
| keywords: | |
| - AI security | |
| - LLM security | |
| - prompt injection | |
| - agentic AI | |
| - red teaming | |
| - hard negatives | |