liu00222/Open-Prompt-Injection

Provides modular factory methods for constructing attacks (combining multiple injection tasks), defenses (DataSentinel detection, PromptLocate localization), and LLM-integrated applications across multiple models (PaLM2, Llama, GPT). Introduces Attack Success Rate (ASV) as a quantitative evaluation metric and includes a detection-plus-localization pipeline that identifies contaminated prompts and recovers original data. Supports configuration-driven experimentation with pre-built task datasets (sentiment analysis, spam detection) and fine-tuned checkpoint integration via LoRA adapters.

406 stars.