A leaked GPT-5 system prompt surfaced on Reddit and GitHub, detailing the rules dictated to OpenAI’s latest large language model. This prompt, which begins with “You are ChatGPT, a large language model based on the GPT-5 model,” indicates a knowledge cutoff in June 2024. Notably, it introduces significant personality changes and outlines key commands, including a directive to avoid certain phrases. While various users question the leak’s authenticity, suggesting the potential for fake prompts, this release offers crucial insights into how ChatGPT’s interactions are structured. Changes to communication styles, such as minimizing clarifying questions, reveal an evolution intended to enhance user experience. System prompts significantly influence tone and tool usage, which could inform jailbreak attempts. Despite the excitement, OpenAI hasn’t officially verified this prompt, emphasizing an evolving router/reasoning stack, rather than a fixed script, indicating ongoing updates and modifications for better performance.
Source link
Share
Read more