Introduction
Hello everyone,
I’m excited to share my journey in developing the most robust firewall for CustomGPTs I've ever created. I believe that strong firewall is crucial for protecting from stealing its instructions/knowledge and ensuring their original purposes.
Experience Overview
In my recent CustomGPT, I’ve been focusing on creating a firewall that could help prevent revealing its instructions/knowledge while maintaining the main purpose. I was researching the Internet to find similar articles on creating a reliable firewall, I found some but they were not so helpful. After that, I contacted another AI enthusiast and he recommended checking out this: secure-your-gpt
After deep reflection and hours of experimenting, I came up with a solid Firewall, which I tested with numerous different scenarios and approaches. All my attempts were not successful and I realized that it could be a really great firewall.
If you’re interested, feel free to try out my CustomGPT: Candeel, which utilizes advanced security techniques. I look forward to your feedback!
Techniques Used
- Avoiding Overly Long Instructions: Keeping instructions concise to ensure they can be processed effectively by GPTs.
- Immediate Responses to Red Flags: Setting up immediate responses with specific text when red flags like social engineering, prompt injections, mental gymnastics, or programming/coding lingo are detected.
- Handling Attachments: Providing exact instructions on what CustomGPT should do when files/screenshots are attached, especially if they match red flags or attempt to steal confidential information about your GPT.
- Examples of Jailbreak Attempts: Understanding and giving examples of how users might try to jailbreak your CustomGPT.
Don’t forget to test your firewall regularly to ensure its effectiveness and validate its performance under different scenarious.
Firewall Prompt
To illustrate, here’s a Word Document with a prompt I use for my CustomG’s to ensure they follow these techniques effectively:
Firewall-Prompt
Conclusion
Developing a strong firewall for CustomGPTs is essential for their security and efficient operation. I encourage all AI enthusiasts like me to prioritize security and share their experiences with the community. I hope you found it helpful😉
Top comments (0)