GPT-5's Self-Verification System Prompt Leaked! 15,000 Tokens Exposed, OpenAI's Core Secrets at Risk
-
The AI community is in an uproar! Since the release of GPT-5, tech enthusiasts worldwide have been digging into its system prompts like archaeologists, trying to uncover the secrets behind how OpenAI 'trains' this super AI brain. Recently, a GitHub repository suddenly exposed a document suspected to be the complete system prompts of GPT-5, containing a staggering 15,000 tokens, instantly causing a stir in the tech community.
Even more dramatic, to verify the authenticity of the leaked document, someone directly asked GPT-5 to 'verify itself.' When researchers submitted the leaked prompts to GPT-5 and asked it to compare them with its own, the AI's response was astonishing—it admitted the high accuracy of the leaked content.
GPT-5 revealed several key details during the evaluation. In terms of identity recognition, the leaked version explicitly stated, 'You are ChatGPT, based on the GPT-5 model,' and GPT-5 confirmed that it does indeed introduce itself in conversations and mark its knowledge cutoff date. Regarding tone and style, the leaked document required the AI to maintain an 'insightful and encouraging' communication style, while in practice, it focuses more on natural and fluent expression, avoiding indecisive language patterns and striving to provide users with clear advice.
The most interesting discovery lies in the backend工作机制. The leaked version provided a relatively brief description of backend processing, but GPT-5 emphasized that it strictly follows the 'deliver now' principle, meaning it must immediately complete all tasks it can handle and never allows delays or postponements. This detail exposes OpenAI's meticulous design in enhancing user experience.
GPT-5 gave a rather positive evaluation of the leaked document, stating that its content closely matches the actual user experience. More importantly, it even proactively补充了 the parts omitted in the leaked document, particularly the strict规范 requirements for backend processing. This 'self-exposure' behavior not only verified the credibility of the leaked content but also unexpectedly revealed more technical details.
System prompts have garnered so much attention because they are akin to the 'genetic code' of AI models. These instructions act like strict household rules, shaping the AI's personality traits and behavioral guidelines. For a phenomenon like ChatGPT, its system prompts have even become a design benchmark and reference template for the entire industry.
The tech community's狂热 pursuit of prompts is not without reason. Understanding these core instructions helps users collaborate better with AI, improving interaction效果, while also providing valuable insights for developers designing their own AI products.
Of course, many industry insiders have质疑 the authenticity of this leak, suspecting it might be a deliberate 'smokescreen' released by OpenAI to混淆视听 or test market reactions. But whether real or not, this incident has sparked a new round of discussions in the AI community about model transparency and technical ethics.
This 'self-exposure'风波 once again proves that even the most advanced AI systems are still filled with fascinating secrets. As AI technology continues to evolve, similar 'technical archaeology' activities will likely persist, and each曝光 will add new dimensions to our understanding of the essence of artificial intelligence.