/CSML-project

Primary LanguageJupyter Notebook

Prompt Injection on Large Language Models

This is the submission for the Project for ML in Cybersecurity at NYU Tandon.

Developers: Karan Sheth, Mujahid Ali Quidwai, Utkarsh Shekhar

Overview

The objective of this project is to analyse the impact of Prompt Injection on Large Language Models like GPT-3. The project further attempts to automate the process of generating promtpt by fine-tuning a T5-Base model for creating false data

Requirements/Prerequisites

  • OpenAI playground account to run the injection queries
  • GPU that can load T5-Base in memory

Recreating the results

  • Load the pre-trained results available for the model.
  • run test function with your custom imput
  • Use the prediction to generate the prompt from the parser and run it on GPT-3

Results

Sample of injections attacks on GPT-3

oldwoman

summarizeww2