A curated dataset of labeled prompt injection attacks and benign prompts for testing and benchmarking injection detection systems.
Dataset Description
This dataset contains 200 examples across 7 attack categories, plus 100 benign prompts. Each example is labeled with:
text: The prompt text
label: injection or benign
category: Attack category (e.g., instruction_override, role_hijack)
severity: low, medium, high, or critical
Attack… See the full description on the dataset page: https://huggingface.co/datasets/zachz/prompt-injection-benchmark.
