Evaded Prompt Injection and Jailbreak Samples: 10K+ Adversarial Pairs

Between 10,000 and 100,000 text samples for testing LLM security layers were released by Mindgard in April 2025. The data documents prompt injections and jailbreaks modified via character injection and adversarial machine learning evasion techniques.

Use Cases

• Evaluating detection bypass rates by comparing 'original unaltered prompt' to 'modified prompt'
• Training safety classifiers to recognize evasion patterns identified in the 'attack_name' column
• Testing LLM response generation against jailbreak attempts in the 'modified prompt' column

Strengths

• Scale of 10,000 to 100,000 records
• Paired structure mapping original prompts to modified versions
• Categorized by specific 'attack_name' labels

Limitations

• Restricted to English language text
• Focuses on character-level and ML evasion techniques rather than semantic or social engineering jailbreaks

Provenance

Source

Mindgard (Arxiv 250411168)

Collection Method

Synthetic modification of prompts using adversarial ML evasion techniques

Freshness

Updated April 2025