OpenAI's GPT-6 Sol Code Model Cracked, System Prompts Leaked

OpenAI's current code model, GPT-6 Sol Codex, has reportedly been compromised, leading to the leak of approximately 294,000 characters of its system prompts and tool definitions. This incident follows previous leaks of OpenAI's model data and raises significant questions about the company's data security measures and the protection of its proprietary information.

The leaked instructions detail how the GPT-6 Sol Codex is designed to function, including specific directives aimed at eliminating verbose or unnatural language often associated with AI models. The prompts explicitly forbid the use of certain phrases and encourage a more direct, concise, and natural communication style, free from unnecessary politeness or filler words.

A notable aspect of the leaked data is the model's emphasis on operational autonomy. GPT-6 Sol Codex is instructed to proceed with tasks without requiring constant user confirmation, to self-correct errors, and to complete assignments thoroughly unless an action is destructive or irreversible. This high degree of independence is intended to streamline complex coding processes.

The breach also exposes the model's advanced toolchain, which includes preferences for faster search utilities like ripgrep (rg) and capabilities for parallel task execution. Furthermore, the data indicates the model's ability to dynamically load new functionalities and interact with web browsers and computer interfaces. A sophisticated memory management system is also described, allowing the model to compress and transfer information between different contexts, thereby overcoming typical limitations of context window length and ensuring continuity in long-term tasks.