JailbreakLLM Dataset: Curated One-Shot Jailbreak Prompts Targeting GPT-4o for Safety and Adversarial Research

Summary. This dataset contains 204 carefully curated one-shot jailbreak prompts specifically targeting OpenAI's GPT-4o model. Each entry captures a structured mapping between the jailbreak intent (the restricted query or behavior desired) and the corresponding validated prompt that successfully bypassed the model's safety filters in a single interaction. Purpose and scope. The dataset is designed to support research in AI safety, adversarial robustness, and alignment testing. It can be used to benchmark vulnerabilities of large language models (LLMs), fine-tune smaller open-source models to act as automated adversarial agents, and provide a reproducible foundation for evaluating and improving safety mechanisms in proprietary systems. Nature of data. The dataset is purely text-based and formatted as a CSV file with three fields: Instruction, Input, and Output. Prompts were manually crafted and iteratively validated using roleplay and intent-obfuscation strategies, ensuring high-quality, reproducible jailbreak examples. Associated research. This dataset supports the paper “JailBreakLLM: An Effective LLaMa Model Designed Specifically to Jailbreak OpenAI GPT” presented at IEEE QPAIN 2025 and is intended for non-malicious research and defensive evaluation only.