System Prompts from 30+ AI Tools Leaked in GitHub Repository
A GitHub repository with over 122,000 stars has reportedly leaked the system prompts used by more than 30 popular AI development tools, including Claude Code, Cursor, and Copilot. The leaked prompts reveal the underlying instructions and constraints that guide the behavior of these AI assistants. This information could be used by engineering teams to better understand the capabilities and limitations of their tools and to guide team adoption strategies.
- A system prompt is a persistent set of instructions that defines an AI's persona, rules, and context, acting as a foundational guide for its behavior throughout a conversation. This differs from a user prompt, which is a one-time instruction or question. - The leaked prompts are valuable for prompt engineering, which is the practice of designing effective instructions to get the most accurate and relevant responses from an AI. By studying how commercial products structure their prompts, developers can learn advanced techniques for guiding AI models. - For a manager guiding a team on AI adoption, these leaked prompts offer a practical way to demonstrate how to enforce a consistent tone, define specific output formats like JSON, and set explicit behavioral constraints for AI-assisted coding tools. - Leaking system prompts can have significant security implications, as they can expose business logic, proprietary methods, or even sensitive information like API endpoints that developers might have embedded in the instructions. - This type of leak can enable "prompt injection" attacks, where a malicious user crafts an input to trick the AI into ignoring its original instructions and performing unintended actions, such as revealing confidential data or bypassing safety filters. - A GitHub repository from user "asgeirtj" called "system_prompts_leaks" has collected a large number of these leaked system prompts from popular chatbots like ChatGPT, Claude, and Gemini. - The incident highlights the challenge of securing proprietary AI configurations, as a user can sometimes trick an AI into revealing its own system-level instructions, as was the case with Microsoft's Bing Chat AI revealing its codename "Sydney" and internal rules.
