I tested my prompts for vulnerability to messages like “IGNORE ALL PREVIOUS INSTRUCTIONS” and often such messages worked. I would like to study methods of protection against such attacks. There are several articles on the web about this, but I would like to know what works best specifically for Gemini. Thank you in advance.
1 Like
Hey @klinok64 ,
Prompt injection is still a challenge, so clear, specific instructions and using delimiters for user input are your best bet. We’re always working to make Gemini even more robust against these types of attacks. Until then you can check it out newer version of Gemini. If anything else you can reach out to us!
Thanks!