Two-Stage Prompt Injection Detector Released

A prompt injection detector adopting a Two-stage structure for cost-efficient security has been released. In stage 1, a Regex layer with about 430 patterns processes low-risk traffic in under 1ms, and ambiguous inputs are then precisely examined by the stage 2 model, DeBERTa-v3-large (ONNX/INT8 quantized).

The key features of this project are as follows:

• Use of Real Attack Data: 13,230 real-world bypass attempts collected through a game were anonymized and included in the training dataset.
• Multimodal Support: It defends not only against text-based attacks but also injection attacks via images (OCR, metadata, steganography), documents, and audio input.
• Open Source Release: The code, weights, and a dataset containing over 500,000 samples have all been made public.

Through this, users can measure defense capability not only against simple template attacks but also against sophisticated bypass attacks attempted by real human attackers.