Graph Neural Networks (GNNs) are powerful tools used in various applications, from social media analysis to drug discovery. But like any powerful tool, they can be vulnerable. Researchers have discovered a new type of attack called a "Graph Injection Attack" (GIA), where malicious actors inject bogus nodes into the graph, disrupting its function. Think of it like introducing a rumor into a social network or a fake citation into a research database. This research focuses on text-attributed graphs (TAGs), where each node has text associated with it, like a post on social media or an academic paper. Previous attacks focused on injecting manipulated embeddings, which are numerical representations of the text. This new research explores injecting the actual text itself, making the attack more realistic and harder to detect. The researchers developed three attack methods. The first, ITGIA, inverts existing attack embeddings into text. While effective, the generated text is often gibberish, making it easily detectable. The second, VTGIA, uses Large Language Models (LLMs) to generate convincing but less effective attack text. The third, WTGIA, strikes a balance. It uses word frequency information to guide LLMs in generating both harmful and coherent text. However, even WTGIA faces challenges. Defenders can easily adapt by using different text embedding methods or even employing LLMs themselves for defense. This research highlights a crucial cat-and-mouse game in AI security. As attackers develop more sophisticated methods, defenders must also innovate to protect the integrity of their systems. The exploration of text-level attacks opens a new frontier in understanding and mitigating vulnerabilities in GNNs, paving the way for more robust and secure AI systems in the future.