Jailbreak and Prompt Injection Dataset for AI Security

Necent's AI Security Dataset contains 691,331 examples for safety research, combining 15 source datasets across 26 languages. It was uploaded to Hugging Face on March 23, 2026.

Use Cases

• Train a classifier to predict the `is_dangerous` label from the `prompt` text and `prompt_type`.
• Analyze the distribution of attack strategies by examining the `prompt_type` and `category` columns across different `model_name` targets.
• Benchmark model robustness by testing responses in the `response` column against known jailbreak `prompt` examples.
• Develop multilingual detection systems using the `prompt` text from 26 different languages.

Strengths

• Large-scale collection of 691,331 adversarial examples.
• Unified data from 15 distinct source datasets, providing variety.
• Multilingual coverage across 26 languages.

Limitations

• The specific row counts, column definitions, and sample data are unavailable for detailed validation.
• Potential class imbalance or label noise is unknown without access to the `is_dangerous` and `category` distributions.
• Dataset freshness is unclear as the source data collection time range is not specified.

Provenance

Source

Hugging Face user 'Necent'; original authoritative sources for the 15 constituent datasets are not detailed.

Collection Method

Combined and unified from 15 existing source datasets focused on AI safety.

Freshness

Last updated on the platform on March 23, 2026.