Researchers have made significant strides in developing large language models (LLMs) to generate high-quality figures from textual descriptions, specifically targeting the creation of TikZ programs that can be rendered as scientific images. A key obstacle in this pursuit is the lack of substantial datasets for Text-to-TikZ, which has hindered the advancement of modeling approaches. To address this, a new methodology, dubbed TikZilla, has been proposed, leveraging reinforcement learning to enhance the quality and scalability of Text-to-TikZ generation1. By harnessing the power of reinforcement learning, TikZilla aims to overcome the limitations of existing datasets and modeling techniques, paving the way for more accurate and efficient figure generation. This breakthrough has significant implications for the security landscape, as LLM developments powered by reinforcement learning introduce new capability and risk surfaces. So what matters to practitioners is that these advancements in LLMs, such as TikZilla, necessitate a thorough examination of their security implications to mitigate potential risks.