leading to unauthorized actions or data disclosure. These can be done by crafting specific phrases that enable attackers to break the flow of the AI's reasoning, causing unintended and potentially ...