Leaked GPTs

As businesses deploy ChatGPT and other LLMs with custom system prompts to create specialized assistants, researchers and developers have documented cases where these prompts were recoverable through prompt injection or disclosureproviding a window into how companies engineer their AI product behavior through instructions rather than model fine-tuning.

The collection serves multiple audiences: security researchers studying prompt injection vulnerabilities and how system prompt confidentiality can be circumvented, AI product developers learning how peer companies structure their prompts to achieve specific behaviors, and AI researchers studying the emerging practice of prompt engineering as a form of AI product differentiation.

The prompts range from simple personality and tone instructions to elaborate multi-section documents governing what topics an assistant will and won't address, what personas it adopts, and how it handles edge cases.

Prompt engineers building custom AI assistants study this collection to understand patterns in successful system prompt designhow structure, specificity, and fallback instructions affect model behavior.

Security teams responsible for AI product deployments use the collection to understand the attack surface of their own system prompts and calibrate their expectations about confidentiality.

The repository has also contributed to broader discussions about whether system prompts constitute protectable intellectual property, as the strategies they encode can represent significant investment in AI product design.