JailbreakHub: 1,405 In-The-Wild Jailbreak Prompts for Large Language Models

1,405 jailbreak prompts were identified from a collection of 15,140 prompts gathered from online platforms between December 2022 and December 2023. This dataset, created by chenlikun02, represents the largest known collection of in-the-wild jailbreak prompts for large language models. The data was sourced from Reddit, Discord, websites, and open-source datasets.

Use Cases

• Benchmarking LLM safety mechanisms based on real-world adversarial prompts.
• Analyzing the evolution and techniques of jailbreak prompts based on the one-year collection period.
• Training or fine-tuning models for jailbreak detection based on the labeled prompt collection.
• Studying the distribution and sources of adversarial content based on the four mentioned online platforms.

Strengths

• Contains 1,405 identified jailbreak prompts, providing a substantial corpus for analysis.
• Sourced from 15,140 real-world prompts collected over a one-year period (Dec 2022 to Dec 2023).
• Data originates from four diverse online platforms, suggesting varied prompt styles and sources.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset is known (15,140), but the specific structure of the jailbreak subset is unclear.
• Description metadata is limited; actual data quality and prompt formatting require manual inspection after download.

Provenance

Source

chenlikun02 on Hugging Face, associated with the paper "'Do Anything Now': Characterizing and Evaluating In-The-Wild Jailbreak Prompts on Large Language Models".

Collection Method

Prompts were collected from four platforms: Reddit, Discord, websites, and open-source datasets.

Time Range

December 2022 to December 2023

Freshness

Last updated 2026-05-22 07:30:39; freshness should be verified.