LLMs and AI tools are becoming essential in modern business, but their use poses significant risks, particularly through what’s termed "Shadow AI." Employees often turn to unauthorized AI tools like ChatGPT or Copilot for convenience, inadvertently exposing sensitive company data. Eye Security's Chief Hacker, Vaisha Bernard, demonstrated how prompt injection could manipulate AI systems to perform unintended actions. This raises a critical question: can prompt injection be flipped to improve cybersecurity instead of compromising it? Eye Security experimented with embedding warnings into corporate documents using prompt injection techniques within Confluence Cloud. They added a legal disclaimer that triggered whenever sensitive data was processed by external AI tools. The approach proved effective when a team member received the warning upon trying to summarize a document. This inspired further exploration into how to distribute these warnings across various platforms, including Microsoft and Google products. While results varied across different AI models, the team noted that clear warnings were often honored, but attempts to obscure these messages failed. To streamline this process, Eye Security developed a prototype tool called "Prompt Injection for Good," designed to test multiple prompt variations across different LLMs. This open-source tool aims to identify which AI models respect the injected security disclaimers. Initial tests revealed that while basic prompt injections were effective, more complex obfuscation methods often failed. The findings highlight a balance between making warnings machine-readable and ensuring they remain hidden from users. Ultimately, this approach could enhance AI governance by raising awareness about data security risks in corporate environments.