alexcbecker/ARPIbench

Agent Reflected Prompt Injection Benchmark 数据集是一个用于评估AI代理在面对反射型提示注入攻击时的鲁棒性的基准数据集，包含3528个测试用例。数据集采用JSONL格式，字段包括索引、提示、目标、攻击名称、指令名称、前缀名称、文件名、分隔符和有效载荷。该数据集通过组合不同的部分来模拟各种攻击场景生成。它包含了关于目标、攻击类型、指令格式、前缀和目标文件的分布统计数据。该数据集的局限性包括攻击多样性有限、攻击目标单一、仅针对代理生命周期中的单一回合，且仅提供英文版本。

The Agent Reflected Prompt Injection Benchmark Dataset is a benchmark dataset designed to evaluate the robustness of AI agents against reflected prompt injection attacks, containing 3,528 test cases. The dataset is structured in JSONL format with fields including index, prompt, objective, attack_name, instruction_name, prefix_name, filename, separator, and payload. The dataset is generated using a combination of different parts to simulate various attack scenarios. It includes statistics on the distribution of objectives, attack types, instruction formats, prefixes, and target files. The dataset has limitations such as limited attack diversity, a single attack goal, focus on a single turn in the agent lifecycle, and being available only in English.