Like many others, my feed was also quickly taken over by the excitement around ChatGPT. I must confess, I am also excited about this, and I have a few ideas of my own. Ideas of integrations abound – let’s create a bot that automatically summarizes twitter threads. Let’s create a bot to summarize whatsapp convesations. Let’s create a bot that will answer questions in conversations, and so on and so on.
With that, many people also showed various “hacks” – getting ChatGPT to respond in an unaligned manner, or printing out its original prompt that OpenAI gave it, etc.
The implications are obvious, without some proper anti-injection techniques, the output from GPT integrations may be unsafe, especially if the output is piped into a scripting enabled environment.