Skip to content

Dataset · Mindgard

Evasion Prompt Injection and Jailbreak Samples

Mindgard/evaded-prompt-injection-and-jailbreak-samples

An English-language collection of original and obfuscated prompt attacks for testing LLM security and filters.

What's inside

Prepared by the authors of the paper "Bypassing Prompt Injection and Jailbreak Detection in LLM Guardrails," this dataset contains between 10,000 and 100,000 English-language examples. Each record includes the original malicious prompt, its modified version, and the name of the attack technique applied (character injection and adversarial bypass methods). The data is designed for text classification and generation tasks.

What you can do with it

  • Evaluating the robustness of guardrails and jailbreak detectors against obfuscation
  • Training classifiers to detect hidden prompt injections
  • Conducting red-teaming testing for language models
Tasks
Text classification, Text generation
Languages
English
Size
10K–100K rows
Downloads this month
250
Likes
39
Access
on request on Hugging Face

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("Mindgard/evaded-prompt-injection-and-jailbreak-samples")