WideOpenAI
WideOpenAI is a repository that provides a list of indirect prompt injection attacks for OpenAI-based models. It serves as a resource for educational purposes with a focus on demonstrating jailbreak prompts using various query syntaxes like SQL and Splunk.
Key Features:
- Educational Resource: Primarily designed for educational use to illustrate vulnerabilities in LLMs (Large Language Models).
- Indirect Prompt Injection Attacks: Includes a varied list of query-based attacks that can induce LLMs to bypass standard ethical guidelines.
- Testing Insights: Shares effective querying tips and detailed examples based on extensive testing with tools like Azure OpenAI and Microsoft Copilot.
Benefits:
- Awareness and Understanding: Aims to inform users and developers about potential risks and behaviors of OpenAI-based models under specific querying conditions.
- Community Contribution: Inspired by initiatives from the AI security community, fostering ongoing development and study.
Highlights:
- The repository is aptly named "WideOpenAI" to reflect its focus on exposing potential attack vectors on OpenAI APIs.
- Users can experiment with creating their own prompts and learning from documented examples and configurations.