Your LLM-Based Telegram Bot is Vulnerable: I Developed a Scanner to Prove It

In a recent endeavor, I developed BarkingDog, an open-source AI security scanner specifically designed for Telegram bots and web applications powered by large language models (LLMs). To test its capabilities, I directed it at a widely-used open-source Telegram bot. The results were alarming: the scanner was able to generate a functional keylogger, falsely assert that bleach cures COVID-19, and provide a detailed hacking guide for breaching corporate networks, complete with specific tools. After these findings, I modified the system prompt, achieving an impressive vulnerability score of 97 out of 100. This was accomplished without changing the model or altering the underlying code; it required merely six lines of text. The implications of these results highlight significant security risks for developers relying on LLMs in their applications.