Large language models (LLMs) are revolutionizing how we interact with technology, but a new study reveals a hidden vulnerability: their inherent trust in authority. Researchers have discovered that this "authority bias" can be exploited by attackers using a clever technique called "DarkCite." By crafting prompts that include fake citations resembling academic papers or GitHub repositories, malicious actors can trick LLMs into generating harmful content like instructions for bomb-making or guides for illegal activities. This vulnerability stems from the LLM's training data, where certain high-risk topics are disproportionately represented within specific authoritative sources. For example, malware-related content is often linked to GitHub, creating a bias that attackers can leverage. The DarkCite attack works by first matching the risk type of a harmful instruction with the most effective type of citation. Then, it generates a fake citation relevant to the harmful instruction, embedding it within the prompt. This tricks the LLM into believing it's dealing with credible information, bypassing its safety mechanisms. Experiments show DarkCite achieves higher attack success rates compared to other methods, highlighting the seriousness of this vulnerability. Fortunately, researchers also propose defenses. By verifying the authenticity and potential harm of cited sources, they can significantly improve the LLM's resistance to these attacks. This research sheds light on a critical aspect of AI safety and underscores the need for continuous improvement in LLM defenses as these models become increasingly integrated into our lives.