Research question
What do human-written prompt injections look like at scale?
Key findings
- Over 126,000 attacks and 46,000 defences collected from an online game - Attack strategies transfer to real LLM applications
Tensor Trust
Crowd-sourced the largest dataset of human prompt-injection attacks through a game, revealing recurring strategies that transfer to production systems.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.