english-net
A prompt-injection detector whose neurons are English yes/no questions, answered by TypeSafe's Jev in one call (79 ms, 2e-5 USD per text). It ties a strong single-question baseline on unseen data (AUC 0.983), and one added sentence raised a missed attack family from 40.8% to 70.…
- Author
- isaacbitran
- Category
- guardrails
- Primitives
- noul
- Source
- github
Classified by jev-1.13.0. Description taken from the source, not generated.