Summary

  • ChatGPT and other Large Language Models (LLMs) are controlled entirely by the text input or prompt given to them by the user.
  • However, if poorly designed, these prompts are susceptible to “prompt injection,” which is akin to SQL injection but with malicious inputs crafted to manipulate the model’s output.
  • The implications of prompt injection include application breakage, data leaks, and exposure of system flaws.
  • Here is a now-removed example of a prompt injection vulnerability in the command line tool ChatGPT.
  • Developers must take caution to mitigate against the risk of prompt injection by practicing strong development and security practices.
  • Educate yourself on best practices by reading the original medium article referenced in this short piece.

By Manish Shivanandhan

Original Article