Type: url
Original: https://github.com/liu00222/Open-Prompt-Injection
Status: generated
Summary
The Open-Prompt-Injection repository, an open-source toolkit and benchmark for prompt injection attacks and defenses in LLM-integrated applications and agents. It works by pairing a target task with an injected task and measuring the attack-success value, with models, tasks, attackers, applications and evaluators assembled through factory functions and JSON config files, and a declarative experiment matrix driving the full experiment set from the associated paper. It also ships two defense components distributed as fine-tuned checkpoints — DataSentinel for detecting injected prompts and PromptLocate for localizing and recovering them — which the repository presents as composable into a detection-then-localization pipeline.